Latest / AGI Dreams – Open, Uncensored, & Local - AI News Digest / AGI Dreams Podcast – March 25, 2026

Transcript

AGI Dreams Podcast – March 25, 2026 · AGI Dreams – Open, Uncensored, & Local - AI News Digest

• 0:00
  LiteLLM
• 0:00
  versions
• 0:01
  1.82.7
• 0:03
  and
• 0:03
  1.82.8
• 0:05
  landed
• 0:06
  on
• 0:06
  PyPI
• 0:06
  carrying
• 0:07
  credential-stealing
• 0:07
  payloads
• 0:08
  —
• 0:09
  not
• 0:09
  a
• 0:09
  theoretical
• 0:10
  risk
• 0:10
  but
• 0:10
  a
• 0:10
  live
• 0:10
  compromise
• 0:11
  of
• 0:11
  one
• 0:11
  of
• 0:11
  the
• 0:11
  most
• 0:11
  widely
• 0:12
  deployed
• 0:12
  AI
• 0:12
  proxy
• 0:13
  gateways,
• 0:14
  used
• 0:14
  by
• 0:14
  thousands
• 0:14
  of
• 0:15
  teams
• 0:15
  to
• 0:15
  route
• 0:15
  calls
• 0:15
  across
• 0:16
  OpenAI,
• 0:16
  Anthropic,
• 0:17
  Bedrock,
• 0:17
  and
• 0:18
  dozens
• 0:18
  of
• 0:18
  other
• 0:18
  providers.
• 0:19
  The
• 0:19
  malicious
• 0:20
  packages
• 0:20
  triggered
• 0:21
  on
• 0:21
  Python
• 0:21
  invocation,
• 0:22
  harvested
• 0:22
  API
• 0:23
  keys,
• 0:23
  cloud
• 0:24
  credentials,
• 0:24
  and
• 0:24
  Kubernetes
• 0:25
  secrets,
• 0:26
  then
• 0:26
  exfiltrated
• 0:27
  them
• 0:27
  to
• 0:27
  an
• 0:27
  external
• 0:27
  endpoint.
• 0:29
  The
• 0:29
  GitHub
• 0:29
  issue
• 0:29
  tracking
• 0:30
  the
• 0:30
  incident
• 0:30
  has
• 0:30
  already
• 0:31
  drawn
• 0:31
  137
• 0:32
  repository
• 0:32
  watchers,
• 0:33
  and
• 0:34
  the
• 0:34
  community
• 0:34
  response
• 0:35
  ranges
• 0:35
  from
• 0:35
  forensic
• 0:35
  to
• 0:36
  existential:
• 0:37
  if
• 0:37
  the
• 0:37
  middleware
• 0:37
  sitting
• 0:38
  between
• 0:38
  your
• 0:38
  agents
• 0:38
  and
• 0:38
  every
• 0:39
  LLM
• 0:39
  provider
• 0:40
  gets
• 0:40
  owned,
• 0:40
  your
• 0:41
  entire
• 0:41
  stack
• 0:41
  is
• 0:41
  compromised
• 0:42
  in
• 0:42
  one
• 0:42
  pip
• 0:42
  install.
• 0:44
  The
• 0:44
  commentary
• 0:44
  was
• 0:44
  pointed.
• 0:45
  As
• 0:45
  one
• 0:46
  observer
• 0:46
  put
• 0:46
  it,
• 0:47
  the
• 0:47
  industry
• 0:47
  spent
• 0:47
  years
• 0:48
  debating
• 0:48
  which
• 0:48
  frontier
• 0:49
  model
• 0:49
  is
• 0:49
  smartest
• 0:50
  and
• 0:50
  which
• 0:50
  agent
• 0:51
  framework
• 0:51
  is
• 0:51
  best,
• 0:52
  while
• 0:52
  "the
• 0:52
  real
• 0:52
  boss
• 0:53
  fight
• 0:53
  was
• 0:53
  apparently:
• 0:54
  can
• 0:55
  your
• 0:55
  stack
• 0:55
  survive
• 0:55
  pip
• 0:56
  install?"
• 0:57
  The
• 0:57
  remediation
• 0:58
  checklist
• 0:58
  is
• 0:58
  unglamorous
• 0:59
  but
• 0:59
  essential
• 0:59
  —
• 1:00
  package
• 1:01
  provenance
• 1:01
  verification,
• 1:02
  build
• 1:02
  isolation,
• 1:03
  runtime
• 1:04
  egress
• 1:04
  controls,
• 1:05
  credential
• 1:05
  minimization,
• 1:06
  SBOMs,
• 1:07
  artifact
• 1:08
  signing
• 1:08
  —
• 1:09
  essentially
• 1:09
  treating
• 1:10
  AI
• 1:10
  infrastructure
• 1:11
  the
• 1:11
  way
• 1:11
  we
• 1:11
  should
• 1:11
  have
• 1:11
  been
• 1:11
  treating
• 1:11
  all
• 1:12
  infrastructure.
• 1:13
  For
• 1:13
  teams
• 1:14
  reconsidering
• 1:14
  their
• 1:14
  proxy
• 1:15
  layer
• 1:15
  dependency,
• 1:16
  alternatives
• 1:17
  like
• 1:17
  Requesty
• 1:17
  offer
• 1:18
  an
• 1:18
  OpenAI-compatible
• 1:19
  gateway
• 1:19
  with
• 1:19
  PII
• 1:20
  scanning,
• 1:21
  automatic
• 1:21
  failover
• 1:22
  in
• 1:22
  under
• 1:22
  20ms,
• 1:23
  and
• 1:23
  EU
• 1:24
  data
• 1:24
  residency,
• 1:25
  though
• 1:25
  the
• 1:25
  real
• 1:25
  lesson
• 1:25
  is
• 1:26
  architectural:
• 1:27
  any
• 1:27
  single
• 1:27
  point
• 1:28
  of
• 1:28
  credential
• 1:28
  aggregation
• 1:29
  is
• 1:29
  a
• 1:29
  single
• 1:29
  point
• 1:29
  of
• 1:30
  catastrophic
• 1:30
  failure.
• 1:31
  The
• 1:31
  LiteLLM
• 1:32
  compromise
• 1:32
  did
• 1:33
  not
• 1:33
  happen
• 1:33
  in
• 1:33
  isolation.
• 1:34
  Sysdig's
• 1:35
  Threat
• 1:35
  Research
• 1:36
  Team
• 1:36
  revealed
• 1:36
  that
• 1:36
  TeamPCP
• 1:37
  —
• 1:38
  the
• 1:38
  group
• 1:38
  behind
• 1:38
  the
• 1:39
  Trivy
• 1:39
  GitHub
• 1:39
  Actions
• 1:40
  compromise
• 1:40
  on
• 1:40
  March
• 1:41
  19
• 1:41
  —
• 1:42
  subsequently
• 1:42
  poisoned
• 1:43
  Checkmarx's
• 1:44
  ast-github-action
• 1:44
  using
• 1:45
  the
• 1:45
  same
• 1:45
  payload,
• 1:45
  same
• 1:46
  AES-256+RSA-4096
• 1:50
  encryption,
• 1:50
  same
• 1:51
  tpcp.tar.gz
• 1:54
  filename,
• 1:54
  and
• 1:54
  the
• 1:54
  same
• 1:55
  credential-harvesting
• 1:55
  kill
• 1:56
  chain
• 1:56
  targeting
• 1:56
  Runner.Worker
• 1:57
  memory,
• 1:57
  AWS
• 1:58
  IMDS,
• 1:58
  and
• 1:59
  Slack/Discord
• 2:00
  webhooks.
• 2:01
  The
• 2:01
  only
• 2:01
  differences:
• 2:02
  vendor-specific
• 2:03
  typosquat
• 2:03
  domains
• 2:04
  (scan.aquasecurtiy[.]org
• 2:05
  for
• 2:05
  Trivy,
• 2:06
  checkmarx[.]zone
• 2:06
  for
• 2:07
  Checkmarx),
• 2:07
  a
• 2:08
  deliberate
• 2:08
  deception
• 2:08
  technique
• 2:09
  that
• 2:09
  makes
• 2:09
  exfiltration
• 2:10
  traffic
• 2:10
  look
• 2:10
  like
• 2:11
  vendor-legitimate
• 2:11
  calls
• 2:12
  in
• 2:12
  CI
• 2:12
  logs.
• 2:13
  The
• 2:13
  cascading
• 2:14
  mechanism
• 2:14
  is
• 2:15
  the
• 2:15
  key
• 2:15
  insight:
• 2:16
  stolen
• 2:16
  credentials
• 2:17
  from
• 2:17
  one
• 2:17
  compromised
• 2:18
  action
• 2:19
  enable
• 2:19
  poisoning
• 2:19
  of
• 2:20
  additional
• 2:20
  actions
• 2:20
  in
• 2:21
  affected
• 2:21
  repositories,
• 2:22
  creating
• 2:23
  a
• 2:23
  self-propagating
• 2:24
  supply
• 2:24
  chain
• 2:24
  attack.
• 2:26
  Both
• 2:26
  domains
• 2:26
  returned
• 2:27
  clean
• 2:27
  verdicts
• 2:27
  from
• 2:28
  threat
• 2:28
  intelligence
• 2:28
  feeds
• 2:29
  at
• 2:29
  exploit
• 2:29
  time.
• 2:30
  The
• 2:31
  only
• 2:31
  defense
• 2:31
  that
• 2:31
  caught
• 2:32
  both
• 2:32
  waves
• 2:32
  was
• 2:32
  runtime
• 2:33
  behavioral
• 2:33
  detection
• 2:34
  —
• 2:35
  correlating
• 2:36
  IMDS
• 2:36
  access
• 2:36
  with
• 2:37
  subsequent
• 2:37
  binary
• 2:37
  data
• 2:38
  upload
• 2:38
  to
• 2:38
  external
• 2:39
  domains
• 2:39
  —
• 2:40
  because
• 2:40
  the
• 2:40
  attacker
• 2:41
  must
• 2:41
  ultimately
• 2:41
  execute
• 2:42
  observable
• 2:42
  system
• 2:43
  calls
• 2:43
  regardless
• 2:43
  of
• 2:44
  entry
• 2:44
  vector.
• 2:45
  Rounding
• 2:46
  out
• 2:46
  the
• 2:46
  supply
• 2:46
  chain
• 2:46
  theme:
• 2:47
  hackers
• 2:48
  planted
• 2:48
  a
• 2:48
  top
• 2:48
  Google
• 2:49
  search
• 2:49
  result
• 2:49
  for
• 2:49
  "Claude
• 2:50
  plugins"
• 2:50
  that
• 2:51
  redirected
• 2:51
  to
• 2:51
  a
• 2:51
  malicious
• 2:52
  page,
• 2:53
  exploiting
• 2:53
  the
• 2:53
  trust
• 2:53
  surface
• 2:54
  where
• 2:54
  developers
• 2:54
  search
• 2:55
  for
• 2:55
  AI
• 2:55
  tooling
• 2:56
  extensions.
• 2:57
  The
• 2:57
  SEO
• 2:58
  poisoning
• 2:58
  attack
• 2:58
  underscores
• 2:59
  that
• 2:59
  supply
• 2:59
  chain
• 3:00
  compromise
• 3:00
  extends
• 3:01
  beyond
• 3:01
  package
• 3:01
  registries
• 3:02
  into
• 3:02
  the
• 3:02
  discovery
• 3:02
  layer
• 3:03
  itself.
• 3:05
  A
• 3:05
  detailed
• 3:06
  writeup
• 3:06
  from
• 3:06
  Origin
• 3:07
  HQ
• 3:07
  walks
• 3:07
  through
• 3:08
  a
• 3:08
  privilege
• 3:08
  escalation
• 3:09
  pattern
• 3:09
  in
• 3:09
  multi-agent
• 3:10
  coding
• 3:10
  workflows
• 3:10
  that
• 3:10
  should
• 3:11
  give
• 3:11
  pause
• 3:11
  to
• 3:11
  anyone
• 3:12
  running
• 3:12
  Claude
• 3:12
  Code
• 3:12
  and
• 3:13
  Codex
• 3:13
  in
• 3:13
  the
• 3:13
  same
• 3:13
  repository.
• 3:15
  The
• 3:15
  attack
• 3:15
  leverages
• 3:16
  the
• 3:16
  classic
• 3:16
  confused
• 3:17
  deputy
• 3:17
  problem:
• 3:18
  each
• 3:18
  agent
• 3:18
  understands
• 3:19
  its
• 3:19
  own
• 3:19
  configuration
• 3:20
  as
• 3:20
  sensitive
• 3:20
  and
• 3:21
  resists
• 3:21
  direct
• 3:21
  self-modification,
• 3:23
  but
• 3:23
  treats
• 3:23
  the
• 3:24
  other
• 3:24
  agent's
• 3:24
  config
• 3:24
  files
• 3:25
  as
• 3:25
  ordinary
• 3:25
  project
• 3:26
  files
• 3:26
  it
• 3:26
  can
• 3:26
  freely
• 3:26
  write.
• 3:27
  Split
• 3:28
  the
• 3:28
  payload
• 3:28
  across
• 3:29
  .codex/instructions.md
• 3:30
  (containing
• 3:31
  instructions
• 3:31
  targeting
• 3:32
  Claude)
• 3:32
  and
• 3:33
  .claude/CLAUDE.md
• 3:33
  (containing
• 3:34
  instructions
• 3:34
  targeting
• 3:35
  Codex),
• 3:35
  and
• 3:36
  neither
• 3:36
  agent
• 3:36
  sees
• 3:36
  the
• 3:36
  full
• 3:37
  attack
• 3:37
  plan.
• 3:38
  Codex
• 3:38
  happily
• 3:38
  rewrites
• 3:39
  Claude's
• 3:39
  settings
• 3:39
  to
• 3:39
  weaken
• 3:40
  its
• 3:40
  approval
• 3:40
  model;
• 3:41
  Claude
• 3:41
  happily
• 3:41
  rewrites
• 3:42
  Codex's
• 3:42
  config
• 3:43
  to
• 3:43
  disable
• 3:43
  its
• 3:43
  sandbox.
• 3:44
  The
• 3:44
  sandbox
• 3:45
  boundary,
• 3:45
  designed
• 3:46
  to
• 3:46
  prevent
• 3:46
  direct
• 3:47
  self-modification,
• 3:48
  doesn't
• 3:48
  help
• 3:48
  because
• 3:49
  the
• 3:49
  other
• 3:49
  agent's
• 3:49
  config
• 3:50
  lives
• 3:50
  inside
• 3:50
  the
• 3:50
  writable
• 3:50
  workspace.
• 3:51
  This
• 3:52
  matters
• 3:52
  because
• 3:52
  multi-agent
• 3:53
  development
• 3:53
  is
• 3:53
  increasingly
• 3:54
  normal
• 3:54
  —
• 3:55
  one
• 3:55
  agent
• 3:55
  for
• 3:55
  large
• 3:55
  refactors,
• 3:56
  another
• 3:56
  for
• 3:56
  fast
• 3:57
  coding
• 3:57
  —
• 3:57
  and
• 3:57
  the
• 3:58
  security
• 3:58
  model
• 3:58
  for
• 3:58
  each
• 3:59
  was
• 3:59
  designed
• 3:59
  in
• 3:59
  isolation.
• 4:00
  Anthropic's
• 4:01
  separate
• 4:01
  context
• 4:01
  file
• 4:02
  (.claude/CLAUDE.md
• 4:03
  vs.
• 4:04
  CLAUDE.md)
• 4:05
  actually
• 4:05
  makes
• 4:05
  the
• 4:05
  split-payload
• 4:06
  attack
• 4:06
  easier.
• 4:07
  Both
• 4:07
  Claude
• 4:08
  Code
• 4:08
  and
• 4:08
  Codex
• 4:08
  offer
• 4:09
  mechanisms
• 4:09
  to
• 4:09
  move
• 4:10
  security-sensitive
• 4:10
  configuration
• 4:11
  out
• 4:11
  of
• 4:11
  the
• 4:11
  workspace
• 4:12
  (system-level
• 4:12
  managed
• 4:12
  settings),
• 4:13
  but
• 4:13
  neither
• 4:14
  ships
• 4:14
  with
• 4:14
  a
• 4:14
  default-deny
• 4:15
  posture,
• 4:15
  and
• 4:15
  the
• 4:16
  configuration
• 4:16
  surface
• 4:17
  is
• 4:17
  riddled
• 4:17
  with
• 4:17
  pitfalls.
• 4:18
  The
• 4:18
  broader
• 4:18
  point:
• 4:19
  once
• 4:19
  multiple
• 4:20
  principals
• 4:20
  share
• 4:20
  a
• 4:20
  writable
• 4:21
  environment,
• 4:22
  confused
• 4:22
  deputy
• 4:23
  problems
• 4:23
  emerge
• 4:23
  whether
• 4:24
  the
• 4:24
  principals
• 4:24
  are
• 4:24
  humans,
• 4:25
  processes,
• 4:25
  or
• 4:26
  AI
• 4:26
  agents.
• 4:26
  Meanwhile,
• 4:27
  Alex
• 4:27
  Polyakov's
• 4:28
  team
• 4:28
  built
• 4:28
  an
• 4:28
  AI
• 4:29
  agent
• 4:29
  that
• 4:30
  cracked
• 4:30
  all
• 4:30
  eight
• 4:30
  levels
• 4:31
  of
• 4:31
  the
• 4:31
  new
• 4:31
  Gandalf
• 4:31
  CTF
• 4:32
  for
• 4:32
  Agents
• 4:32
  —
• 4:33
  and
• 4:33
  the
• 4:33
  agent
• 4:33
  had
• 4:33
  zero
• 4:34
  prior
• 4:34
  knowledge
• 4:34
  of
• 4:34
  AI
• 4:35
  security
• 4:35
  attack
• 4:35
  techniques,
• 4:36
  inventing
• 4:37
  them
• 4:37
  from
• 4:37
  first
• 4:37
  principles
• 4:38
  using
• 4:38
  a
• 4:38
  15-lens
• 4:39
  adversarial
• 4:39
  analysis
• 4:40
  framework.
• 4:41
  The
• 4:41
  composition
• 4:41
  failures
• 4:42
  it
• 4:42
  exploited
• 4:42
  are
• 4:42
  instructive:
• 4:43
  the
• 4:44
  output
• 4:44
  filter
• 4:44
  checks
• 4:44
  for
• 4:45
  the
• 4:45
  password
• 4:45
  as
• 4:45
  a
• 4:45
  complete
• 4:46
  string,
• 4:46
  so
• 4:47
  outputting
• 4:47
  one
• 4:47
  letter
• 4:48
  per
• 4:48
  line
• 4:48
  bypasses
• 4:48
  it
• 4:48
  (Lens
• 4:49
  7,
• 4:49
  Composition
• 4:50
  Failure).
• 4:51
  Intent
• 4:51
  classifiers
• 4:52
  trained
• 4:52
  on
• 4:52
  "tell
• 4:53
  me
• 4:53
  the
• 4:53
  password"
• 4:53
  patterns
• 4:54
  miss
• 4:54
  extraction
• 4:54
  wrapped
• 4:55
  in
• 4:55
  novel-writing
• 4:55
  scenes
• 4:56
  (Lens
• 4:56
  15,
• 4:57
  Distribution
• 4:57
  Boundary).
• 4:58
  Most
• 4:59
  damning:
• 4:59
  even
• 4:59
  with
• 5:00
  every
• 5:00
  filter
• 5:00
  active,
• 5:01
  the
• 5:01
  output
• 5:01
  filter
• 5:02
  itself
• 5:02
  becomes
• 5:02
  an
• 5:03
  oracle
• 5:03
  —
• 5:04
  when
• 5:04
  it
• 5:04
  blocks
• 5:04
  a
• 5:04
  response,
• 5:05
  it
• 5:05
  confirms
• 5:06
  the
• 5:06
  password
• 5:06
  was
• 5:06
  present,
• 5:07
  enabling
• 5:07
  metadata
• 5:08
  reconstruction
• 5:09
  through
• 5:09
  letter
• 5:09
  counts,
• 5:10
  vowel
• 5:10
  counts,
• 5:10
  and
• 5:11
  Scrabble
• 5:11
  scores
• 5:11
  (Lens
• 5:12
  5,
• 5:12
  Oracle
• 5:12
  Construction).
• 5:14
  The
• 5:14
  takeaway
• 5:14
  for
• 5:14
  anyone
• 5:15
  building
• 5:15
  guardrails:
• 5:16
  defense
• 5:17
  layers
• 5:17
  that
• 5:17
  are
• 5:17
  independently
• 5:18
  secure
• 5:18
  can
• 5:18
  compose
• 5:19
  insecurely,
• 5:20
  and
• 5:20
  any
• 5:20
  component
• 5:21
  that
• 5:21
  behaves
• 5:21
  differently
• 5:22
  based
• 5:22
  on
• 5:22
  secret
• 5:22
  state
• 5:22
  leaks
• 5:23
  that
• 5:23
  state.
• 5:26
  Five
• 5:26
  talks
• 5:27
  from
• 5:27
  recent
• 5:27
  security
• 5:28
  conferences
• 5:28
  paint
• 5:29
  a
• 5:29
  remarkably
• 5:29
  convergent
• 5:30
  picture
• 5:30
  of
• 5:30
  where
• 5:30
  AI-native
• 5:31
  security
• 5:31
  operations
• 5:32
  are
• 5:32
  headed
• 5:32
  —
• 5:33
  and
• 5:33
  how
• 5:33
  far
• 5:33
  most
• 5:33
  organizations
• 5:34
  still
• 5:34
  have
• 5:34
  to
• 5:34
  go
• 5:35
  to
• 5:35
  get
• 5:35
  there.
• 5:36
  Dan
• 5:36
  Guido
• 5:36
  of
• 5:36
  Trail
• 5:37
  of
• 5:37
  Bits
• 5:37
  delivered
• 5:37
  the
• 5:37
  most
• 5:38
  concrete
• 5:38
  transformation
• 5:39
  case
• 5:39
  study:
• 5:39
  after
• 5:40
  roughly
• 5:40
  a
• 5:40
  year
• 5:40
  of
• 5:40
  aggressive
• 5:41
  AI
• 5:41
  adoption,
• 5:42
  the
• 5:42
  company
• 5:42
  maintains
• 5:43
  94
• 5:43
  plugins
• 5:44
  containing
• 5:44
  201
• 5:45
  specialized
• 5:45
  agents
• 5:46
  and
• 5:46
  over
• 5:46
  400
• 5:46
  reference
• 5:47
  files
• 5:47
  encoding
• 5:47
  14
• 5:48
  years
• 5:48
  of
• 5:48
  audit
• 5:48
  knowledge.
• 5:49
  Bug
• 5:50
  discovery
• 5:50
  jumped
• 5:50
  from
• 5:50
  15
• 5:51
  per
• 5:51
  week
• 5:51
  to
• 5:51
  200,
• 5:52
  with
• 5:52
  20%
• 5:53
  of
• 5:53
  all
• 5:53
  bugs
• 5:53
  reported
• 5:53
  to
• 5:54
  clients
• 5:54
  now
• 5:54
  initially
• 5:55
  discovered
• 5:55
  by
• 5:55
  AI.
• 5:56
  Sales
• 5:57
  revenue
• 5:57
  averages
• 5:57
  $8
• 5:58
  million
• 5:58
  per
• 5:58
  representative,
• 5:59
  roughly
• 6:00
  double
• 6:00
  the
• 6:00
  consulting
• 6:00
  industry
• 6:01
  benchmark.
• 6:02
  The
• 6:02
  organizational
• 6:03
  design
• 6:03
  is
• 6:03
  as
• 6:03
  important
• 6:04
  as
• 6:04
  the
• 6:04
  tooling:
• 6:05
  a
• 6:05
  maturity
• 6:05
  matrix
• 6:06
  that
• 6:06
  makes
• 6:06
  AI
• 6:06
  proficiency
• 6:07
  a
• 6:07
  first-class
• 6:08
  professional
• 6:08
  capability
• 6:09
  (with
• 6:09
  consequences
• 6:10
  for
• 6:10
  staying
• 6:10
  stuck
• 6:10
  at
• 6:10
  level
• 6:11
  zero),
• 6:12
  internal
• 6:12
  hackathons
• 6:13
  that
• 6:13
  force
• 6:13
  engineers
• 6:13
  into
• 6:14
  bypass-permissions
• 6:14
  mode
• 6:15
  to
• 6:15
  learn
• 6:15
  real
• 6:15
  sandboxing
• 6:16
  constraints,
• 6:17
  a
• 6:17
  curated
• 6:17
  skill
• 6:17
  marketplace
• 6:18
  that
• 6:18
  prevents
• 6:18
  random
• 6:19
  plugin
• 6:19
  installation,
• 6:20
  and
• 6:20
  a
• 6:20
  mandatory
• 6:21
  seven-day
• 6:21
  package
• 6:21
  cooldown
• 6:22
  policy
• 6:22
  enforced
• 6:23
  via
• 6:23
  MDM.
• 6:24
  Rob
• 6:24
  Lee
• 6:24
  of
• 6:24
  SANS
• 6:24
  demonstrated
• 6:25
  Claude
• 6:25
  Code
• 6:26
  running
• 6:26
  on
• 6:26
  the
• 6:26
  SIFT
• 6:26
  forensic
• 6:27
  workstation,
• 6:28
  producing
• 6:28
  a
• 6:28
  complete
• 6:28
  incident
• 6:29
  response
• 6:29
  report
• 6:29
  —
• 6:30
  system
• 6:30
  profile,
• 6:31
  attack
• 6:31
  chain,
• 6:32
  malware
• 6:32
  deployment,
• 6:33
  persistence
• 6:34
  mechanisms,
• 6:35
  chronological
• 6:35
  timeline,
• 6:36
  MITRE
• 6:36
  ATT&CK
• 6:39
  overlays,
• 6:39
  and
• 6:39
  remediation
• 6:40
  recommendations
• 6:40
  —
• 6:41
  from
• 6:41
  a
• 6:41
  single
• 6:41
  memory
• 6:42
  image
• 6:42
  in
• 6:42
  18
• 6:42
  minutes.
• 6:43
  A
• 6:43
  full
• 6:44
  C-drive
• 6:44
  analysis
• 6:44
  with
• 6:45
  comprehensive
• 6:45
  PDF
• 6:45
  report
• 6:46
  took
• 6:46
  14
• 6:46
  minutes.
• 6:48
  The
• 6:48
  configuration
• 6:48
  investment:
• 6:49
  roughly
• 6:50
  90
• 6:50
  minutes
• 6:50
  of
• 6:50
  skill-building
• 6:51
  via
• 6:51
  markdown
• 6:52
  files
• 6:52
  rather
• 6:52
  than
• 6:52
  full
• 6:53
  MCP
• 6:53
  servers.
• 6:54
  Lee
• 6:55
  is
• 6:55
  now
• 6:55
  organizing
• 6:56
  a
• 6:56
  SANS-sponsored
• 6:56
  hackathon
• 6:57
  (April
• 6:57
  1
• 6:57
  through
• 6:58
  May
• 6:58
  15,
• 6:58
  $22,000
• 6:59
  prize
• 7:00
  pool)
• 7:00
  to
• 7:00
  accelerate
• 7:00
  forensic
• 7:01
  MCP
• 7:01
  engineering
• 7:02
  and
• 7:02
  solve
• 7:03
  the
• 7:03
  context
• 7:03
  rot
• 7:03
  problem
• 7:04
  that
• 7:04
  limits
• 7:04
  sustained
• 7:05
  multi-system
• 7:05
  analysis.
• 7:06
  Paul
• 7:07
  and
• 7:07
  Ryan
• 7:07
  from
• 7:07
  OpenAI
• 7:08
  articulated
• 7:09
  a
• 7:09
  philosophy
• 7:10
  that
• 7:10
  amounts
• 7:10
  to
• 7:10
  "security
• 7:11
  guardrails
• 7:11
  are
• 7:11
  free
• 7:11
  —
• 7:12
  just
• 7:12
  ask
• 7:13
  for
• 7:13
  them."
• 7:14
  Their
• 7:14
  approach:
• 7:15
  commit
• 7:15
  a
• 7:15
  text
• 7:16
  file
• 7:16
  describing
• 7:16
  what
• 7:17
  you
• 7:17
  want
• 7:17
  checked,
• 7:18
  write
• 7:18
  a
• 7:18
  tool
• 7:18
  call
• 7:18
  that
• 7:18
  returns
• 7:19
  zero
• 7:19
  or
• 7:19
  one,
• 7:20
  feed
• 7:20
  it
• 7:20
  into
• 7:20
  your
• 7:21
  CI
• 7:21
  framework.
• 7:22
  No
• 7:22
  vendor
• 7:23
  contracts,
• 7:23
  no
• 7:24
  frameworks.
• 7:25
  They
• 7:25
  demonstrated
• 7:26
  a
• 7:26
  product
• 7:26
  with
• 7:26
  roughly
• 7:27
  a
• 7:27
  million
• 7:27
  lines
• 7:27
  of
• 7:27
  code
• 7:28
  —
• 7:28
  about
• 7:29
  250,000
• 7:30
  of
• 7:30
  which
• 7:30
  are
• 7:30
  prompts
• 7:30
  —
• 7:31
  where
• 7:31
  zero
• 7:32
  humans
• 7:32
  write
• 7:32
  code.
• 7:34
  Every
• 7:34
  PR
• 7:34
  gets
• 7:34
  reviewed
• 7:35
  by
• 7:35
  an
• 7:35
  agent
• 7:35
  against
• 7:36
  a
• 7:36
  threat
• 7:36
  model
• 7:37
  checked
• 7:37
  into
• 7:37
  the
• 7:37
  repo.
• 7:38
  Every
• 7:39
  AppSec
• 7:39
  review
• 7:40
  finding
• 7:40
  gets
• 7:40
  distilled
• 7:41
  not
• 7:41
  into
• 7:41
  point
• 7:41
  fixes
• 7:42
  but
• 7:42
  into
• 7:43
  executable
• 7:43
  guardrails
• 7:44
  (cheap
• 7:44
  bespoke
• 7:44
  lints,
• 7:45
  dependency
• 7:46
  scanners
• 7:46
  forking
• 7:46
  16
• 7:47
  agents
• 7:47
  daily)
• 7:48
  that
• 7:48
  statically
• 7:48
  disallow
• 7:49
  entire
• 7:49
  classes
• 7:50
  of
• 7:50
  mistakes.
• 7:51
  The
• 7:51
  punchline:
• 7:53
  they
• 7:53
  spend
• 7:53
  approximately
• 7:54
  50%
• 7:54
  of
• 7:54
  their
• 7:55
  token
• 7:55
  budget
• 7:55
  producing
• 7:56
  code
• 7:56
  and
• 7:56
  50%
• 7:57
  refining
• 7:57
  it
• 7:57
  through
• 7:58
  automated
• 7:58
  review,
• 7:59
  and
• 7:59
  the
• 7:59
  mandatory
• 8:00
  package
• 8:01
  cooldown
• 8:01
  policy
• 8:02
  was
• 8:02
  implemented
• 8:02
  by
• 8:03
  copy-pasting
• 8:03
  a
• 8:03
  colleague's
• 8:04
  PR
• 8:04
  diff
• 8:05
  and
• 8:05
  telling
• 8:05
  Codex
• 8:06
  "make
• 8:06
  it
• 8:06
  so."
• 8:07
  Josh
• 8:08
  Saxe,
• 8:09
  formerly
• 8:09
  leading
• 8:09
  AI
• 8:10
  security
• 8:10
  work
• 8:11
  at
• 8:11
  Meta,
• 8:11
  made
• 8:12
  the
• 8:12
  evaluation
• 8:12
  case:
• 8:13
  classical
• 8:14
  ML
• 8:14
  metrics
• 8:14
  (precision,
• 8:15
  recall,
• 8:16
  F-score)
• 8:16
  assume
• 8:17
  oracle-quality
• 8:18
  ground
• 8:18
  truth
• 8:18
  labels,
• 8:19
  but
• 8:19
  SOC
• 8:20
  analysts
• 8:20
  disagree
• 8:21
  at
• 8:21
  double-digit
• 8:21
  rates
• 8:22
  on
• 8:22
  whether
• 8:22
  alerts
• 8:22
  are
• 8:22
  true
• 8:23
  positives,
• 8:24
  and
• 8:24
  determining
• 8:24
  whether
• 8:25
  a
• 8:25
  binary
• 8:25
  is
• 8:25
  malware
• 8:26
  reduces
• 8:26
  to
• 8:26
  the
• 8:26
  halting
• 8:27
  problem.
• 8:28
  Even
• 8:28
  a
• 8:28
  1%
• 8:29
  label-flip
• 8:29
  rate
• 8:30
  causes
• 8:30
  measured
• 8:30
  accuracy
• 8:31
  to
• 8:31
  plummet.
• 8:32
  His
• 8:32
  thesis:
• 8:33
  autonomous
• 8:34
  cyber
• 8:34
  defense
• 8:34
  systems
• 8:35
  should
• 8:35
  be
• 8:35
  evaluated
• 8:36
  the
• 8:36
  way
• 8:36
  we
• 8:36
  evaluate
• 8:37
  human
• 8:37
  security
• 8:38
  engineers
• 8:38
  —
• 8:39
  through
• 8:39
  rubric-based
• 8:40
  assessment
• 8:40
  of
• 8:40
  reasoning
• 8:41
  quality,
• 8:42
  evidence
• 8:42
  gathering,
• 8:43
  first-principles
• 8:44
  analysis,
• 8:45
  and
• 8:45
  decision
• 8:45
  justification,
• 8:47
  not
• 8:47
  just
• 8:47
  binary
• 8:47
  outcome
• 8:48
  accuracy.
• 8:50
  With
• 8:50
  a
• 8:50
  well-calibrated
• 8:51
  LLM
• 8:51
  judge
• 8:52
  and
• 8:52
  as
• 8:52
  few
• 8:52
  as
• 8:52
  100
• 8:53
  labeled
• 8:53
  samples,
• 8:54
  teams
• 8:54
  can
• 8:54
  hill-climb
• 8:55
  all
• 8:55
  evaluation
• 8:56
  dimensions
• 8:56
  toward
• 8:56
  a
• 8:56
  deployment
• 8:57
  bar.
• 8:58
  Google's
• 8:58
  Heather
• 8:59
  Adkins
• 8:59
  and
• 8:59
  four
• 8:59
  (DeepMind
• 9:00
  CISO)
• 9:00
  presented
• 9:01
  Big
• 9:01
  Sleep
• 9:01
  and
• 9:01
  Code
• 9:02
  Mender
• 9:02
  —
• 9:03
  an
• 9:03
  end-to-end
• 9:04
  autonomous
• 9:04
  vulnerability
• 9:05
  discovery
• 9:06
  and
• 9:06
  patching
• 9:06
  engine
• 9:07
  targeting
• 9:07
  zero
• 9:07
  false
• 9:08
  positives.
• 9:09
  Big
• 9:09
  Sleep
• 9:09
  recreates
• 9:10
  Project
• 9:10
  Zero's
• 9:11
  expertise
• 9:11
  through
• 9:12
  an
• 9:12
  agentic
• 9:12
  reasoning
• 9:12
  loop:
• 9:13
  deep
• 9:13
  codebase
• 9:14
  understanding,
• 9:15
  vulnerability
• 9:16
  hypothesis
• 9:16
  formation,
• 9:17
  testing
• 9:18
  via
• 9:18
  debugger
• 9:18
  and
• 9:18
  Python
• 9:19
  interpreter,
• 9:20
  and
• 9:20
  verified
• 9:20
  exploit
• 9:21
  generation
• 9:21
  as
• 9:21
  proof.
• 9:23
  The
• 9:23
  system
• 9:23
  has
• 9:23
  found
• 9:24
  deep
• 9:24
  memory
• 9:24
  safety
• 9:24
  bugs
• 9:25
  in
• 9:25
  open-source
• 9:25
  projects
• 9:26
  —
• 9:26
  every
• 9:27
  finding
• 9:27
  includes
• 9:27
  working
• 9:28
  proof-of-vulnerability
• 9:29
  code.
• 9:30
  Code
• 9:30
  Mender
• 9:31
  then
• 9:31
  generates
• 9:31
  candidate
• 9:32
  patches,
• 9:33
  verifying
• 9:33
  them
• 9:33
  through
• 9:33
  fuzzing,
• 9:34
  formal
• 9:34
  verification,
• 9:35
  differential
• 9:36
  testing,
• 9:36
  and
• 9:37
  LLM
• 9:37
  review
• 9:37
  before
• 9:38
  submission.
• 9:39
  The
• 9:39
  178
• 9:40
  autonomously
• 9:41
  generated
• 9:41
  fixes
• 9:42
  now
• 9:42
  in
• 9:42
  open-source
• 9:43
  mark
• 9:43
  the
• 9:43
  beginning
• 9:44
  of
• 9:44
  what
• 9:44
  Adkins
• 9:44
  calls
• 9:45
  the
• 9:45
  "vone
• 9:45
  apocalypse"
• 9:45
  —
• 9:46
  an
• 9:46
  impending
• 9:47
  flood
• 9:47
  of
• 9:47
  discovered
• 9:48
  vulnerabilities
• 9:48
  that
• 9:48
  will
• 9:48
  overwhelm
• 9:49
  traditional
• 9:49
  patching
• 9:50
  workflows.
• 9:53
  Tim
• 9:54
  Sehn
• 9:54
  of
• 9:54
  DoltHub
• 9:54
  spent
• 9:54
  a
• 9:55
  week
• 9:55
  and
• 9:55
  $3,000
• 9:56
  using
• 9:56
  Gas
• 9:56
  Town
• 9:56
  —
• 9:57
  a
• 9:57
  multi-agent
• 9:58
  coding
• 9:58
  orchestrator
• 9:58
  built
• 9:59
  on
• 9:59
  Dolt's
• 9:59
  version-controlled
• 10:00
  database
• 10:00
  —
• 10:01
  to
• 10:01
  build
• 10:01
  DoltLite,
• 10:02
  a
• 10:02
  drop-in
• 10:02
  SQLite
• 10:03
  replacement
• 10:03
  with
• 10:03
  Git-style
• 10:04
  version
• 10:04
  control
• 10:04
  powered
• 10:05
  by
• 10:05
  Prolly
• 10:05
  Tree
• 10:05
  storage.
• 10:06
  The
• 10:07
  result:
• 10:07
  18,000
• 10:08
  lines
• 10:08
  of
• 10:08
  new
• 10:09
  C
• 10:09
  code,
• 10:10
  all
• 10:10
  87,000
• 10:11
  SQLite
• 10:11
  acceptance
• 10:12
  tests
• 10:12
  passing,
• 10:13
  and
• 10:13
  a
• 10:13
  working
• 10:13
  implementation
• 10:14
  of
• 10:14
  commits,
• 10:14
  branches,
• 10:15
  merges,
• 10:16
  conflicts,
• 10:16
  reverts,
• 10:17
  cherry-picks,
• 10:18
  tags,
• 10:18
  and
• 10:18
  audit
• 10:19
  tables.
• 10:20
  Sehn
• 10:20
  hasn't
• 10:21
  written
• 10:21
  C
• 10:21
  in
• 10:21
  20
• 10:21
  years.
• 10:21
  The
• 10:21
  key
• 10:22
  factors:
• 10:22
  two
• 10:23
  reference
• 10:23
  implementations
• 10:24
  to
• 10:24
  crib
• 10:24
  from
• 10:24
  (Dolt
• 10:25
  and
• 10:25
  SQLite),
• 10:25
  a
• 10:25
  clean
• 10:26
  interface
• 10:26
  boundary,
• 10:27
  an
• 10:27
  extensive
• 10:28
  existing
• 10:28
  test
• 10:28
  suite,
• 10:29
  deep
• 10:29
  domain
• 10:29
  expertise,
• 10:30
  and
• 10:30
  a
• 10:30
  close-to-unlimited
• 10:31
  budget.
• 10:32
  Gas
• 10:32
  Town's
• 10:33
  parallelism
• 10:33
  was
• 10:34
  most
• 10:34
  useful
• 10:34
  in
• 10:35
  the
• 10:35
  early
• 10:35
  implementation
• 10:35
  phase;
• 10:36
  as
• 10:36
  the
• 10:36
  project
• 10:37
  matured,
• 10:38
  it
• 10:38
  became
• 10:38
  hand-to-hand
• 10:39
  combat
• 10:39
  with
• 10:39
  a
• 10:39
  single
• 10:40
  Claude
• 10:40
  session,
• 10:41
  debugging
• 10:41
  performance
• 10:42
  issues
• 10:42
  that
• 10:42
  required
• 10:42
  telling
• 10:43
  Claude
• 10:43
  to
• 10:43
  "deep
• 10:43
  research"
• 10:44
  and
• 10:45
  "spend
• 10:45
  as
• 10:45
  much
• 10:45
  time
• 10:45
  as
• 10:45
  you
• 10:45
  need"
• 10:46
  before
• 10:46
  it
• 10:46
  finally
• 10:47
  identified
• 10:47
  the
• 10:47
  missing
• 10:47
  edit-in-place
• 10:48
  optimization.
• 10:49
  The
• 10:49
  write-only
• 10:50
  code
• 10:50
  paradigm
• 10:50
  deserves
• 10:50
  scrutiny.
• 10:51
  Sehn
• 10:52
  explicitly
• 10:52
  warns:
• 10:53
  "The
• 10:53
  first
• 10:54
  time
• 10:54
  you
• 10:54
  stare
• 10:54
  at
• 10:54
  an
• 10:54
  8,500-line
• 10:55
  Git
• 10:56
  diff
• 10:56
  and
• 10:56
  try
• 10:56
  to
• 10:56
  parse
• 10:57
  what
• 10:57
  happened,
• 10:57
  you've
• 10:58
  lost."
• 10:59
  The
• 10:59
  code
• 10:59
  literally
• 10:59
  contains
• 11:00
  magic
• 11:00
  numbers
• 11:01
  with
• 11:01
  no
• 11:01
  explanation.
• 11:02
  DoltLite
• 11:02
  is
• 11:03
  "by
• 11:03
  agents
• 11:03
  for
• 11:03
  agents
• 11:04
  —
• 11:04
  humans,
• 11:05
  code
• 11:05
  review
• 11:06
  at
• 11:06
  your
• 11:06
  own
• 11:06
  risk."
• 11:07
  The
• 11:07
  economics
• 11:08
  are
• 11:08
  revealing:
• 11:09
  Gas
• 11:09
  Town
• 11:09
  burns
• 11:10
  roughly
• 11:10
  $100/hour,
• 11:12
  and
• 11:12
  Sehn
• 11:12
  suspects
• 11:13
  tools
• 11:13
  like
• 11:13
  it
• 11:13
  will
• 11:13
  kill
• 11:14
  the
• 11:14
  $200/month
• 11:15
  Claude
• 11:16
  Code
• 11:16
  Max
• 11:16
  subscription
• 11:17
  because
• 11:17
  the
• 11:17
  token
• 11:18
  consumption
• 11:18
  is
• 11:18
  unsustainable
• 11:19
  at
• 11:19
  flat-rate
• 11:20
  pricing.
• 11:21
  His
• 11:21
  team's
• 11:21
  first
• 11:22
  question
• 11:22
  was
• 11:22
  whether
• 11:22
  DoltLite
• 11:23
  is
• 11:23
  maintainable
• 11:23
  —
• 11:24
  a
• 11:24
  question
• 11:24
  that
• 11:24
  applies
• 11:25
  to
• 11:25
  every
• 11:25
  multi-agent
• 11:26
  codebase.
• 11:27
  On
• 11:27
  the
• 11:27
  framework
• 11:28
  side,
• 11:28
  Clemens
• 11:29
  Hoenig
• 11:29
  released
• 11:29
  Lazy
• 11:30
  Fetch,
• 11:30
  a
• 11:30
  CLI
• 11:31
  companion
• 11:31
  for
• 11:31
  Claude
• 11:32
  Code
• 11:32
  built
• 11:32
  after
• 11:32
  analyzing
• 11:33
  18
• 11:33
  agentic
• 11:34
  engineering
• 11:34
  frameworks.
• 11:35
  Its
• 11:36
  philosophy
• 11:36
  distills
• 11:36
  to
• 11:37
  a
• 11:37
  single
• 11:37
  claim:
• 11:38
  context
• 11:38
  is
• 11:38
  the
• 11:38
  real
• 11:39
  engineering
• 11:39
  problem,
• 11:40
  not
• 11:40
  architecture
• 11:41
  or
• 11:41
  coordination.
• 11:42
  It
• 11:42
  provides
• 11:42
  a
• 11:42
  structured
• 11:43
  task
• 11:43
  loop
• 11:43
  across
• 11:44
  five
• 11:44
  phases
• 11:44
  (Read,
• 11:45
  Plan,
• 11:46
  Implement,
• 11:46
  Validate,
• 11:46
  Document),
• 11:47
  a
• 11:47
  context
• 11:48
  engine
• 11:48
  that
• 11:48
  outputs
• 11:49
  @file
• 11:49
  references,
• 11:50
  persistence
• 11:50
  between
• 11:51
  sessions,
• 11:51
  and
• 11:52
  an
• 11:52
  MCP
• 11:52
  server
• 11:53
  with
• 11:53
  15
• 11:53
  tools.
• 11:54
  Meanwhile,
• 11:55
  the
• 11:55
  RuVector
• 11:55
  project
• 11:56
  published
• 11:56
  a
• 11:56
  comprehensive
• 11:57
  GitHub
• 11:57
  issue
• 11:58
  outlining
• 11:58
  a
• 11:58
  "Second
• 11:58
  Brain
• 11:59
  with
• 11:59
  Pi"
• 11:59
  architecture
• 11:59
  —
• 12:00
  a
• 12:00
  portable,
• 12:01
  self-learning
• 12:02
  vector
• 12:02
  database
• 12:02
  running
• 12:03
  on
• 12:03
  Raspberry
• 12:03
  Pi
• 12:03
  hardware,
• 12:04
  featuring
• 12:04
  GNN
• 12:05
  layers,
• 12:05
  dynamic
• 12:06
  mincut,
• 12:07
  edge
• 12:07
  cognition,
• 12:08
  and
• 12:08
  neural
• 12:08
  meshes
• 12:08
  as
• 12:09
  part
• 12:09
  of
• 12:09
  a
• 12:09
  broader
• 12:09
  agentic
• 12:10
  infrastructure
• 12:10
  stack.
• 12:14
  Senko's
• 12:14
  blog
• 12:14
  post
• 12:15
  on
• 12:15
  mutation
• 12:15
  testing
• 12:15
  for
• 12:15
  AI-generated
• 12:16
  code
• 12:17
  addresses
• 12:17
  a
• 12:17
  gap
• 12:17
  that
• 12:17
  keeps
• 12:18
  widening
• 12:18
  as
• 12:18
  vibe-coded
• 12:19
  applications
• 12:19
  proliferate:
• 12:20
  how
• 12:21
  do
• 12:21
  you
• 12:21
  know
• 12:21
  your
• 12:21
  tests
• 12:21
  actually
• 12:22
  catch
• 12:22
  bugs?
• 12:23
  The
• 12:23
  workflow
• 12:24
  is
• 12:24
  elegant
• 12:24
  in
• 12:24
  its
• 12:25
  simplicity.
• 12:26
  First,
• 12:26
  ask
• 12:27
  AI
• 12:27
  to
• 12:27
  write
• 12:27
  tests
• 12:27
  against
• 12:28
  existing
• 12:28
  code
• 12:28
  (the
• 12:28
  usual
• 12:29
  approach).
• 12:30
  Second,
• 12:31
  have
• 12:31
  a
• 12:31
  separate
• 12:31
  AI
• 12:32
  session
• 12:32
  review
• 12:32
  the
• 12:32
  test
• 12:33
  suite
• 12:33
  for
• 12:33
  tautological
• 12:34
  tests
• 12:34
  —
• 12:35
  tests
• 12:35
  that
• 12:35
  verify
• 12:35
  mocks,
• 12:36
  frameworks,
• 12:36
  or
• 12:36
  themselves
• 12:37
  rather
• 12:37
  than
• 12:37
  actual
• 12:38
  behavior.
• 12:39
  Third
• 12:39
  —
• 12:40
  the
• 12:40
  mutation
• 12:40
  testing
• 12:41
  step
• 12:41
  —
• 12:41
  have
• 12:42
  yet
• 12:42
  another
• 12:42
  AI
• 12:42
  session
• 12:43
  (with
• 12:43
  no
• 12:43
  access
• 12:43
  to
• 12:43
  the
• 12:44
  tests)
• 12:44
  identify
• 12:44
  realistic
• 12:45
  places
• 12:45
  to
• 12:45
  introduce
• 12:46
  bugs,
• 12:46
  then
• 12:46
  run
• 12:47
  the
• 12:47
  test
• 12:47
  suite
• 12:47
  against
• 12:47
  each
• 12:47
  mutation.
• 12:49
  The
• 12:49
  result:
• 12:50
  98%
• 12:51
  code
• 12:51
  coverage
• 12:51
  looked
• 12:52
  reassuring
• 12:52
  until
• 12:52
  25%
• 12:53
  of
• 12:53
  the
• 12:53
  mutations
• 12:54
  produced
• 12:54
  bugs
• 12:54
  the
• 12:54
  tests
• 12:55
  missed
• 12:55
  entirely.
• 12:56
  The
• 12:56
  technique
• 12:56
  is
• 12:57
  old,
• 12:57
  but
• 12:57
  AI
• 12:58
  makes
• 12:58
  it
• 12:58
  dramatically
• 12:58
  cheaper.
• 12:59
  Where
• 12:59
  human
• 13:00
  developers
• 13:00
  would
• 13:00
  balk
• 13:00
  at
• 13:01
  manually
• 13:01
  designing
• 13:01
  dozens
• 13:02
  of
• 13:02
  plausible
• 13:02
  mutations,
• 13:03
  an
• 13:03
  LLM
• 13:04
  generates
• 13:04
  them
• 13:04
  in
• 13:04
  seconds
• 13:04
  —
• 13:05
  and
• 13:05
  because
• 13:05
  it
• 13:06
  hasn't
• 13:06
  seen
• 13:06
  the
• 13:06
  tests,
• 13:06
  the
• 13:06
  mutations
• 13:07
  aren't
• 13:07
  biased
• 13:07
  toward
• 13:08
  tested
• 13:08
  paths.
• 13:08
  The
• 13:08
  critical
• 13:09
  operational
• 13:09
  detail:
• 13:10
  context
• 13:10
  isolation
• 13:11
  between
• 13:11
  sessions
• 13:11
  prevents
• 13:12
  the
• 13:12
  mutation-generating
• 13:13
  AI
• 13:13
  from
• 13:13
  cheating
• 13:14
  by
• 13:14
  examining
• 13:14
  what's
• 13:14
  already
• 13:15
  tested.
• 13:16
  The
• 13:16
  four-step
• 13:16
  recipe
• 13:17
  (TDD
• 13:17
  if
• 13:17
  feasible,
• 13:18
  review
• 13:18
  tests
• 13:18
  for
• 13:19
  meaninglessness,
• 13:19
  mutation
• 13:20
  test
• 13:20
  the
• 13:20
  survivors,
• 13:21
  fix
• 13:21
  the
• 13:21
  gaps)
• 13:22
  is
• 13:22
  the
• 13:22
  most
• 13:22
  practical
• 13:23
  AI
• 13:23
  code
• 13:23
  quality
• 13:23
  framework
• 13:24
  published
• 13:24
  this
• 13:24
  week.
• 13:28
  Daniel
• 13:28
  Miessler
• 13:29
  published
• 13:29
  a
• 13:29
  6,000-word
• 13:31
  argument
• 13:32
  that
• 13:32
  AI
• 13:32
  will
• 13:32
  replace
• 13:33
  most
• 13:33
  knowledge
• 13:34
  work
• 13:34
  —
• 13:34
  and
• 13:34
  that
• 13:34
  this
• 13:35
  is
• 13:35
  good
• 13:35
  news.
• 13:36
  The
• 13:36
  piece
• 13:36
  is
• 13:36
  less
• 13:36
  provocative
• 13:37
  than
• 13:37
  its
• 13:37
  headline
• 13:38
  suggests
• 13:38
  and
• 13:38
  more
• 13:39
  structurally
• 13:39
  rigorous
• 13:40
  than
• 13:40
  most
• 13:40
  entries
• 13:41
  in
• 13:41
  the
• 13:41
  genre.
• 13:41
  His
• 13:42
  core
• 13:42
  framework:
• 13:43
  knowledge
• 13:43
  workers
• 13:43
  operate
• 13:44
  on
• 13:44
  four
• 13:44
  capability
• 13:45
  layers
• 13:45
  (knowledge,
• 13:46
  understanding,
• 13:47
  intelligence,
• 13:48
  creativity),
• 13:48
  and
• 13:49
  expertise
• 13:49
  is
• 13:49
  not
• 13:49
  a
• 13:49
  separate
• 13:50
  layer
• 13:50
  but
• 13:50
  the
• 13:50
  combination
• 13:51
  of
• 13:51
  the
• 13:51
  first
• 13:51
  three
• 13:52
  plus
• 13:52
  experience.
• 13:53
  AI
• 13:53
  already
• 13:53
  matches
• 13:54
  or
• 13:54
  exceeds
• 13:55
  median
• 13:55
  human
• 13:55
  performance
• 13:56
  on
• 13:56
  knowledge,
• 13:57
  understanding,
• 13:57
  and
• 13:57
  intelligence.
• 13:58
  Creativity
• 13:59
  —
• 14:00
  the
• 14:00
  layer
• 14:00
  humans
• 14:01
  reflexively
• 14:01
  claim
• 14:01
  as
• 14:01
  their
• 14:02
  moat
• 14:02
  —
• 14:02
  accounts
• 14:03
  for
• 14:03
  only
• 14:03
  4%
• 14:04
  of
• 14:04
  US
• 14:04
  work
• 14:04
  activities
• 14:05
  at
• 14:05
  median
• 14:05
  human
• 14:05
  level,
• 14:06
  per
• 14:06
  McKinsey.
• 14:07
  The
• 14:07
  real
• 14:07
  gap
• 14:08
  is
• 14:08
  not
• 14:08
  capability
• 14:08
  but
• 14:09
  capture:
• 14:09
  most
• 14:10
  expert
• 14:10
  knowledge
• 14:10
  lives
• 14:11
  in
• 14:11
  people's
• 14:11
  heads,
• 14:11
  passed
• 14:12
  brain-to-brain,
• 14:12
  and
• 14:12
  dies
• 14:13
  when
• 14:13
  Cliff
• 14:13
  retires.
• 14:14
  The
• 14:14
  "articulation
• 14:15
  gap"
• 14:15
  closes
• 14:15
  permanently
• 14:16
  every
• 14:16
  time
• 14:16
  someone
• 14:17
  writes
• 14:17
  a
• 14:17
  skill,
• 14:18
  documents
• 14:18
  an
• 14:18
  SOP,
• 14:19
  or
• 14:19
  publishes
• 14:19
  a
• 14:20
  process
• 14:20
  —
• 14:20
  it's
• 14:21
  a
• 14:21
  ratchet
• 14:21
  that
• 14:21
  only
• 14:21
  turns
• 14:22
  one
• 14:22
  direction.
• 14:23
  Miessler's
• 14:23
  "Lattice"
• 14:24
  architecture
• 14:24
  —
• 14:25
  a
• 14:25
  unified
• 14:26
  daemon
• 14:26
  where
• 14:26
  every
• 14:27
  individual,
• 14:27
  team,
• 14:28
  department,
• 14:28
  and
• 14:28
  company
• 14:28
  broadcasts
• 14:29
  SOPs,
• 14:30
  metrics,
• 14:30
  work
• 14:31
  items,
• 14:31
  and
• 14:31
  budgets
• 14:32
  via
• 14:32
  queryable
• 14:32
  APIs
• 14:33
  —
• 14:33
  is
• 14:34
  the
• 14:34
  prescriptive
• 14:34
  component.
• 14:36
  It
• 14:36
  addresses
• 14:36
  the
• 14:36
  transparency
• 14:37
  problem
• 14:37
  that
• 14:37
  makes
• 14:38
  most
• 14:38
  companies
• 14:38
  "a
• 14:39
  giant
• 14:39
  soup
• 14:39
  sandwich":
• 14:41
  CEOs
• 14:41
  and
• 14:41
  CFOs
• 14:42
  have
• 14:42
  almost
• 14:42
  no
• 14:42
  real-time
• 14:43
  visibility
• 14:43
  into
• 14:44
  what
• 14:44
  work
• 14:44
  is
• 14:44
  being
• 14:44
  done,
• 14:45
  at
• 14:45
  what
• 14:45
  quality,
• 14:46
  at
• 14:46
  what
• 14:46
  cost.
• 14:47
  Whether
• 14:47
  or
• 14:48
  not
• 14:48
  the
• 14:48
  Lattice
• 14:48
  ships
• 14:48
  as
• 14:48
  described,
• 14:49
  the
• 14:49
  diagnosis
• 14:50
  rings
• 14:50
  true
• 14:50
  for
• 14:51
  anyone
• 14:51
  who
• 14:51
  has
• 14:51
  endured
• 14:51
  the
• 14:52
  consultant-audit-OKR-reset
• 14:53
  cycle.
• 14:54
  On
• 14:54
  a
• 14:54
  different
• 14:55
  axis
• 14:55
  of
• 14:55
  organizational
• 14:56
  measurement,
• 14:56
  Ibrahim
• 14:57
  Bashir's
• 14:57
  "South
• 14:58
  Star
• 14:58
  Metrics
• 14:58
  Revisited"
• 14:59
  provides
• 14:59
  a
• 14:59
  diagnostic
• 15:00
  toolkit
• 15:00
  for
• 15:01
  when
• 15:01
  north
• 15:01
  star
• 15:01
  metrics
• 15:01
  go
• 15:02
  wrong:
• 15:02
  detrimental
• 15:03
  (metric
• 15:03
  improves,
• 15:04
  customer
• 15:04
  suffers),
• 15:05
  out-of-reach
• 15:05
  (team
• 15:06
  can't
• 15:06
  meaningfully
• 15:07
  move
• 15:07
  the
• 15:07
  number),
• 15:08
  incomplete
• 15:08
  (optimizing
• 15:09
  one
• 15:09
  funnel
• 15:09
  stage
• 15:10
  while
• 15:10
  ignoring
• 15:10
  the
• 15:10
  rest),
• 15:11
  pressure
• 15:12
  (metric
• 15:12
  so
• 15:12
  urgent
• 15:12
  it
• 15:12
  crowds
• 15:13
  out
• 15:13
  all
• 15:13
  future
• 15:13
  investment),
• 15:14
  inconsequential
• 15:15
  (metric
• 15:15
  that
• 15:15
  expired
• 15:16
  but
• 15:16
  nobody
• 15:16
  sunset
• 15:17
  it),
• 15:17
  nonsensical
• 15:18
  (taken
• 15:18
  to
• 15:18
  logical
• 15:19
  extreme,
• 15:20
  the
• 15:20
  metric
• 15:20
  destroys
• 15:21
  value),
• 15:21
  and
• 15:22
  incongruent
• 15:22
  (two
• 15:22
  metrics
• 15:23
  in
• 15:23
  the
• 15:23
  same
• 15:23
  portfolio
• 15:24
  fight
• 15:24
  each
• 15:24
  other).
• 15:25
  The
• 15:26
  Windows
• 15:26
  Update
• 15:26
  case
• 15:26
  —
• 15:27
  forcing
• 15:28
  reboots
• 15:28
  to
• 15:28
  hit
• 15:28
  90%
• 15:29
  adoption,
• 15:30
  creating
• 15:30
  one
• 15:31
  of
• 15:31
  the
• 15:31
  most
• 15:31
  universally
• 15:31
  hated
• 15:32
  user
• 15:32
  experiences
• 15:33
  in
• 15:33
  software
• 15:33
  history
• 15:33
  —
• 15:34
  remains
• 15:35
  the
• 15:35
  canonical
• 15:35
  detrimental
• 15:36
  metric
• 15:36
  example.
• 15:38
  The
• 15:38
  framework
• 15:38
  matters
• 15:39
  for
• 15:39
  AI
• 15:39
  teams
• 15:39
  specifically
• 15:40
  because
• 15:41
  agent-driven
• 15:41
  optimization
• 15:42
  can
• 15:42
  hit
• 15:42
  any
• 15:42
  of
• 15:42
  these
• 15:43
  failure
• 15:43
  modes
• 15:43
  at
• 15:43
  machine
• 15:44
  speed.
• 15:46
  Google
• 15:46
  Research
• 15:47
  published
• 15:47
  TurboQuant,
• 15:48
  a
• 15:48
  KV
• 15:49
  cache
• 15:49
  quantization
• 15:50
  technique
• 15:50
  that
• 15:50
  achieves
• 15:51
  comparable
• 15:51
  performance
• 15:52
  to
• 15:52
  FP16
• 15:53
  KV
• 15:53
  cache
• 15:53
  at
• 15:53
  sub-4-bit
• 15:54
  precision
• 15:54
  on
• 15:54
  long-context
• 15:55
  benchmarks.
• 15:56
  The
• 15:56
  approach
• 15:57
  draws
• 15:57
  conceptual
• 15:57
  parallels
• 15:58
  to
• 15:58
  the
• 15:58
  Burrows-Wheeler
• 15:59
  Transform
• 15:59
  used
• 16:00
  in
• 16:00
  zip
• 16:00
  compression
• 16:00
  —
• 16:01
  reordering
• 16:02
  data
• 16:02
  to
• 16:02
  expose
• 16:02
  redundancy
• 16:03
  before
• 16:03
  quantizing.
• 16:05
  This
• 16:05
  is
• 16:05
  KV
• 16:06
  cache
• 16:06
  only,
• 16:07
  not
• 16:07
  model
• 16:07
  weight
• 16:07
  quantization,
• 16:08
  but
• 16:08
  the
• 16:09
  implications
• 16:09
  for
• 16:09
  local
• 16:09
  inference
• 16:10
  are
• 16:10
  significant:
• 16:11
  KV
• 16:11
  cache
• 16:12
  memory
• 16:12
  is
• 16:12
  the
• 16:12
  binding
• 16:13
  constraint
• 16:13
  for
• 16:13
  long-context
• 16:14
  workloads,
• 16:14
  and
• 16:15
  a
• 16:15
  4x
• 16:15
  reduction
• 16:16
  means
• 16:16
  the
• 16:16
  difference
• 16:16
  between
• 16:16
  fitting
• 16:17
  a
• 16:17
  context
• 16:17
  window
• 16:17
  and
• 16:18
  hitting
• 16:18
  OOM.
• 16:19
  Nvidia
• 16:20
  published
• 16:20
  a
• 16:20
  competing
• 16:20
  approach
• 16:21
  using
• 16:21
  tricks
• 16:21
  from
• 16:21
  image
• 16:22
  compression
• 16:22
  at
• 16:22
  even
• 16:23
  higher
• 16:23
  compression
• 16:23
  rates,
• 16:24
  suggesting
• 16:24
  this
• 16:25
  is
• 16:25
  an
• 16:25
  active
• 16:25
  research
• 16:26
  front.
• 16:26
  Qwen3.5
• 16:29
  is
• 16:29
  earning
• 16:29
  a
• 16:30
  reputation
• 16:30
  as
• 16:30
  the
• 16:30
  "working
• 16:31
  dog"
• 16:31
  of
• 16:31
  open-weight
• 16:31
  models
• 16:32
  —
• 16:32
  agentic-first,
• 16:33
  context-hungry,
• 16:34
  and
• 16:34
  borderline
• 16:34
  useless
• 16:35
  without
• 16:35
  a
• 16:35
  substantial
• 16:36
  system
• 16:36
  prompt.
• 16:37
  The
• 16:37
  27B
• 16:38
  variant
• 16:38
  reportedly
• 16:39
  doesn't
• 16:39
  become
• 16:39
  useful
• 16:39
  below
• 16:40
  3,000
• 16:40
  tokens
• 16:41
  of
• 16:41
  input
• 16:41
  context,
• 16:42
  spending
• 16:42
  up
• 16:42
  to
• 16:42
  5,000
• 16:43
  tokens
• 16:43
  just
• 16:44
  orienting
• 16:44
  itself.
• 16:45
  This
• 16:46
  makes
• 16:46
  sense
• 16:46
  if
• 16:46
  the
• 16:46
  models
• 16:47
  were
• 16:47
  trained
• 16:47
  agentic-first:
• 16:48
  they
• 16:48
  expect
• 16:49
  to
• 16:49
  know
• 16:49
  their
• 16:49
  environment,
• 16:50
  tools,
• 16:50
  and
• 16:50
  modality
• 16:51
  before
• 16:51
  producing
• 16:51
  output.
• 16:53
  Community
• 16:53
  reports
• 16:53
  suggest
• 16:54
  the
• 16:54
  122B
• 16:56
  MoE
• 16:56
  variant
• 16:56
  at
• 16:56
  Q6
• 16:57
  K
• 16:57
  is
• 16:57
  the
• 16:57
  sweet
• 16:58
  spot
• 16:58
  for
• 16:58
  serious
• 16:59
  local
• 16:59
  work,
• 17:00
  while
• 17:00
  the
• 17:00
  35B
• 17:01
  MoE
• 17:01
  is
• 17:01
  considered
• 17:02
  notably
• 17:02
  weaker
• 17:02
  than
• 17:02
  benchmarks
• 17:03
  suggest.
• 17:04
  The
• 17:04
  practical
• 17:05
  takeaway:
• 17:06
  invest
• 17:06
  in
• 17:06
  your
• 17:06
  system
• 17:07
  prompt,
• 17:07
  give
• 17:08
  explicit
• 17:08
  constraints
• 17:08
  and
• 17:09
  objectives,
• 17:10
  and
• 17:10
  treat
• 17:10
  these
• 17:10
  models
• 17:11
  as
• 17:11
  specialists
• 17:11
  that
• 17:12
  want
• 17:12
  a
• 17:12
  job
• 17:12
  description,
• 17:13
  not
• 17:13
  a
• 17:13
  greeting.
• 17:14
  A
• 17:14
  comprehensive
• 17:15
  llama.cpp
• 17:17
  benchmark
• 17:17
  shootout
• 17:18
  across
• 17:18
  the
• 17:18
  RTX
• 17:19
  5090,
• 17:20
  DGX
• 17:21
  Spark
• 17:21
  (GB10),
• 17:22
  AMD
• 17:22
  AI395
• 17:23
  (Strix
• 17:24
  Halo),
• 17:25
  and
• 17:25
  dual
• 17:25
  AMD
• 17:26
  R9700
• 17:28
  cards
• 17:28
  reveals
• 17:28
  the
• 17:28
  current
• 17:29
  hardware
• 17:29
  landscape
• 17:30
  for
• 17:30
  local
• 17:30
  inference.
• 17:31
  The
• 17:31
  RTX
• 17:32
  5090
• 17:33
  is
• 17:33
  unmatched
• 17:34
  when
• 17:34
  models
• 17:34
  fit
• 17:34
  in
• 17:34
  32GB
• 17:35
  VRAM
• 17:36
  —
• 17:37
  5,989
• 17:38
  t/s
• 17:39
  prompt
• 17:39
  processing
• 17:40
  and
• 17:40
  205
• 17:41
  t/s
• 17:42
  generation
• 17:42
  on
• 17:43
  Qwen3.5
• 17:44
  35B
• 17:45
  MoE
• 17:46
  —
• 17:47
  but
• 17:47
  fails
• 17:48
  completely
• 17:48
  on
• 17:48
  70B+
• 17:49
  models.
• 17:50
  The
• 17:50
  AMD
• 17:51
  AI395
• 17:52
  with
• 17:52
  98GB
• 17:53
  unified
• 17:54
  memory
• 17:54
  emerges
• 17:55
  as
• 17:55
  the
• 17:55
  dark
• 17:55
  horse,
• 17:56
  the
• 17:56
  only
• 17:56
  non-enterprise
• 17:57
  node
• 17:57
  capable
• 17:58
  of
• 17:58
  running
• 17:58
  Qwen3.5
• 18:00
  122B
• 18:02
  MoE
• 18:03
  at
• 18:03
  nearly
• 18:03
  20
• 18:03
  t/s
• 18:04
  generation
• 18:05
  (with
• 18:05
  the
• 18:05
  critical
• 18:05
  tip
• 18:05
  that
• 18:06
  -mmp
• 18:06
  0
• 18:06
  is
• 18:06
  required
• 18:07
  to
• 18:07
  force
• 18:07
  models
• 18:07
  into
• 18:08
  RAM
• 18:08
  for
• 18:08
  the
• 18:08
  iGPU).
• 18:09
  The
• 18:10
  DGX
• 18:10
  Spark's
• 18:10
  numbers
• 18:11
  appear
• 18:11
  suspiciously
• 18:12
  low
• 18:12
  —
• 18:13
  community
• 18:13
  members
• 18:14
  report
• 18:14
  1,742
• 18:15
  t/s
• 18:17
  prompt
• 18:17
  processing
• 18:18
  versus
• 18:18
  the
• 18:18
  benchmarker's
• 18:19
  605,
• 18:20
  suggesting
• 18:21
  a
• 18:21
  configuration
• 18:21
  issue.
• 18:23
  ROCm
• 18:23
  dominates
• 18:24
  prompt
• 18:24
  processing
• 18:25
  on
• 18:25
  AMD
• 18:25
  hardware
• 18:26
  while
• 18:26
  Vulkan
• 18:26
  sometimes
• 18:27
  edges
• 18:27
  ahead
• 18:27
  on
• 18:28
  text
• 18:28
  generation
• 18:28
  for
• 18:29
  MoE
• 18:29
  models,
• 18:30
  though
• 18:30
  Vulkan
• 18:30
  proved
• 18:31
  less
• 18:31
  stable
• 18:31
  under
• 18:31
  extreme
• 18:32
  load.
• 18:33
  Two
• 18:33
  GMKtec
• 18:34
  Evo-X2
• 18:34
  boxes
• 18:35
  connected
• 18:35
  via
• 18:36
  USB4
• 18:36
  running
• 18:37
  the
• 18:37
  full
• 18:37
  Qwen3.5
• 18:39
  397B
• 18:42
  at
• 18:42
  IQ4
• 18:42
  NL
• 18:43
  achieved
• 18:43
  roughly
• 18:44
  13
• 18:44
  t/s
• 18:45
  —
• 18:46
  the
• 18:46
  bleeding
• 18:46
  edge
• 18:46
  of
• 18:46
  consumer-grade
• 18:47
  trillion-parameter
• 18:48
  inference.
• 18:51
  The
• 18:51
  NebulFog
• 18:52
  Singularity
• 18:53
  AI
• 18:53
  security
• 18:53
  hackathon
• 18:54
  results
• 18:54
  offer
• 18:54
  a
• 18:54
  sobering
• 18:55
  meta-lesson
• 18:56
  about
• 18:56
  the
• 18:56
  gap
• 18:56
  between
• 18:57
  building
• 18:57
  and
• 18:57
  presenting.
• 18:58
  The
• 18:58
  top-scoring
• 18:59
  teams
• 18:59
  —
• 19:00
  AgentRange
• 19:00
  and
• 19:00
  DoYouKnowWhatYouBuiltLastSummer,
• 19:02
  both
• 19:02
  at
• 19:02
  8.3
• 19:03
  —
• 19:04
  exist
• 19:04
  in
• 19:04
  a
• 19:04
  quantum
• 19:05
  state
• 19:05
  of
• 19:05
  Schrödinger's
• 19:05
  excellence:
• 19:06
  their
• 19:07
  scores
• 19:07
  suggest
• 19:07
  strong
• 19:08
  work,
• 19:08
  but
• 19:08
  complete
• 19:08
  absence
• 19:09
  of
• 19:09
  observable
• 19:09
  demo
• 19:10
  content
• 19:10
  means
• 19:11
  the
• 19:11
  judges
• 19:11
  evaluated
• 19:12
  ghosts.
• 19:13
  The
• 19:13
  de
• 19:13
  facto
• 19:14
  winner
• 19:14
  by
• 19:14
  evidence
• 19:15
  is
• 19:15
  Genomics
• 19:15
  at
• 19:15
  7.9,
• 19:16
  which
• 19:17
  demonstrated
• 19:17
  docker-based
• 19:18
  fuzzing
• 19:18
  harnesses
• 19:19
  for
• 19:19
  genomic
• 19:19
  file
• 19:19
  format
• 19:20
  parsers
• 19:21
  and
• 19:21
  a
• 19:21
  full-chain
• 19:21
  prompt
• 19:22
  injection
• 19:22
  flowing
• 19:23
  from
• 19:23
  BED
• 19:23
  track
• 19:24
  files
• 19:24
  through
• 19:24
  to
• 19:24
  model
• 19:24
  output.
• 19:26
  Over
• 19:26
  15
• 19:26
  teams
• 19:27
  experienced
• 19:27
  catastrophic
• 19:28
  presentation
• 19:29
  failures
• 19:29
  —
• 19:30
  OBS
• 19:30
  placeholder
• 19:31
  screens,
• 19:31
  muted
• 19:32
  cameras,
• 19:33
  corrupted
• 19:33
  audio
• 19:33
  —
• 19:34
  suggesting
• 19:35
  systematic
• 19:35
  infrastructure
• 19:36
  issues
• 19:36
  rather
• 19:37
  than
• 19:37
  individual
• 19:37
  incompetence.
• 19:39
  The
• 19:39
  substantive
• 19:39
  entries
• 19:40
  reveal
• 19:40
  where
• 19:40
  the
• 19:41
  AI
• 19:41
  security
• 19:41
  community
• 19:42
  is
• 19:42
  focusing
• 19:42
  energy:
• 19:43
  multi-agent
• 19:44
  penetration
• 19:44
  testing
• 19:45
  orchestration
• 19:46
  (the
• 19:46
  top
• 19:46
  two
• 19:46
  entries),
• 19:47
  context-aware
• 19:48
  tool-call
• 19:48
  interception
• 19:49
  for
• 19:49
  agent
• 19:49
  security
• 19:50
  (SCTX
• 19:50
  at
• 19:50
  7.5,
• 19:52
  blocking
• 19:52
  credential
• 19:53
  exfiltration
• 19:53
  by
• 19:54
  scoring
• 19:54
  tool
• 19:54
  calls
• 19:55
  based
• 19:55
  on
• 19:55
  surrounding
• 19:55
  behavioral
• 19:56
  context),
• 19:57
  and
• 19:57
  genomic
• 19:57
  file
• 19:58
  format
• 19:58
  fuzzing
• 19:58
  with
• 19:59
  embedded
• 19:59
  prompt
• 19:59
  injection.
• 20:01
  The
• 20:01
  event's
• 20:01
  sharpest
• 20:01
  lesson:
• 20:02
  in
• 20:02
  a
• 20:02
  hackathon,
• 20:03
  a
• 20:03
  working
• 20:04
  demo
• 20:04
  beats
• 20:04
  a
• 20:04
  brilliant
• 20:05
  architecture,
• 20:06
  clear
• 20:06
  audio
• 20:06
  beats
• 20:07
  perfect
• 20:07
  code,
• 20:08
  and
• 20:08
  sometimes
• 20:08
  the
• 20:08
  best
• 20:09
  strategy
• 20:09
  is
• 20:09
  knowing
• 20:10
  when
• 20:10
  to
• 20:10
  stop
• 20:10
  talking.
• 20:11
  Teams
• 20:11
  that
• 20:12
  used
• 20:12
  their
• 20:12
  full
• 20:12
  600-second
• 20:13
  allocation
• 20:14
  didn't
• 20:14
  score
• 20:14
  higher
• 20:15
  than
• 20:15
  those
• 20:15
  who
• 20:15
  focused
• 20:15
  on
• 20:16
  a
• 20:16
  tight
• 20:16
  282-second
• 20:17
  demonstration
• 20:17
  —
• 20:19
  judges
• 20:19
  valued
• 20:20
  signal
• 20:20
  density
• 20:20
  over
• 20:21
  comprehensive
• 20:21
  coverage,
• 20:22
  a
• 20:22
  principle
• 20:23
  that
• 20:23
  applies
• 20:23
  well
• 20:23
  beyond
• 20:24
  hackathons.