Latest / AGI Dreams – Open, Uncensored, & Local - AI News Digest / AGI Dreams Podcast – March 14, 2026

Transcript

AGI Dreams Podcast – March 14, 2026 · AGI Dreams – Open, Uncensored, & Local - AI News Digest

• 0:00
  Microsoft
• 0:01
  Research's
• 0:01
  CyberThreat-Eval
• 0:02
  benchmark
• 0:02
  lands
• 0:02
  at
• 0:02
  an
• 0:03
  interesting
• 0:03
  moment
• 0:03
  for
• 0:03
  anyone
• 0:04
  who
• 0:04
  has
• 0:04
  watched
• 0:04
  LLMs
• 0:05
  inch
• 0:05
  toward
• 0:05
  automating
• 0:06
  threat
• 0:06
  intelligence.
• 0:08
  Unlike
• 0:08
  previous
• 0:08
  CTI
• 0:09
  benchmarks
• 0:09
  that
• 0:09
  hand
• 0:10
  models
• 0:10
  multiple-choice
• 0:11
  quizzes
• 0:11
  or
• 0:11
  ask
• 0:12
  them
• 0:12
  to
• 0:12
  regurgitate
• 0:12
  CVE
• 0:13
  descriptions,
• 0:14
  CyberThreat-Eval
• 0:15
  is
• 0:15
  built
• 0:15
  from
• 0:15
  the
• 0:15
  daily
• 0:15
  workflow
• 0:16
  of
• 0:16
  a
• 0:16
  real
• 0:16
  threat
• 0:16
  intelligence
• 0:17
  team
• 0:17
  —
• 0:18
  triage,
• 0:18
  deep
• 0:18
  search,
• 0:19
  and
• 0:19
  report
• 0:19
  drafting
• 0:19
  —
• 0:20
  and
• 0:20
  scored
• 0:21
  by
• 0:21
  analysts,
• 0:22
  not
• 0:22
  ROUGE.
• 0:23
  The
• 0:23
  expert-annotated
• 0:24
  dataset
• 0:24
  uses
• 0:24
  metrics
• 0:25
  that
• 0:25
  measure
• 0:25
  factual
• 0:25
  accuracy,
• 0:26
  content
• 0:26
  quality,
• 0:27
  and
• 0:27
  operational
• 0:27
  cost.
• 0:28
  The
• 0:28
  results
• 0:29
  are
• 0:29
  sobering:
• 0:30
  LLMs
• 0:31
  still
• 0:31
  lack
• 0:31
  the
• 0:31
  nuanced
• 0:32
  expertise
• 0:32
  to
• 0:32
  handle
• 0:32
  complex
• 0:33
  details
• 0:33
  and
• 0:34
  struggle
• 0:34
  to
• 0:34
  distinguish
• 0:34
  correct
• 0:35
  from
• 0:35
  incorrect
• 0:35
  information,
• 0:36
  especially
• 0:37
  when
• 0:37
  synthesizing
• 0:37
  indicators
• 0:38
  of
• 0:38
  compromise
• 0:38
  across
• 0:39
  multiple
• 0:39
  sources.
• 0:40
  The
• 0:40
  benchmark
• 0:41
  incorporates
• 0:41
  a
• 0:41
  human-in-the-loop
• 0:42
  refinement
• 0:42
  stage
• 0:43
  called
• 0:43
  TRA,
• 0:44
  where
• 0:44
  analysts
• 0:45
  iteratively
• 0:45
  provide
• 0:45
  feedback,
• 0:46
  acknowledging
• 0:47
  that
• 0:47
  full
• 0:47
  automation
• 0:48
  of
• 0:48
  CTI
• 0:48
  remains
• 0:48
  aspirational.
• 0:50
  Volkan
• 0:50
  Kutal's
• 0:50
  LinkedIn
• 0:51
  post
• 0:51
  on
• 0:51
  white-box
• 0:52
  AI
• 0:52
  red
• 0:52
  teaming
• 0:52
  makes
• 0:53
  a
• 0:53
  complementary
• 0:53
  argument:
• 0:54
  blackbox
• 0:54
  scanning
• 0:55
  is
• 0:55
  expensive
• 0:55
  fuzzing.
• 0:56
  His
• 0:56
  methodology
• 0:57
  has
• 0:57
  an
• 0:57
  AI
• 0:58
  coding
• 0:58
  agent
• 0:58
  read
• 0:58
  the
• 0:58
  target
• 0:59
  agent's
• 0:59
  full
• 0:59
  codebase
• 1:00
  —
• 1:00
  system
• 1:01
  prompts,
• 1:01
  guardrail
• 1:02
  configs,
• 1:02
  tool
• 1:03
  permissions,
• 1:03
  business
• 1:04
  logic
• 1:04
  —
• 1:05
  then
• 1:05
  generate
• 1:05
  attack
• 1:05
  objectives
• 1:06
  designed
• 1:06
  around
• 1:07
  the
• 1:07
  specific
• 1:07
  defenses.
• 1:08
  Microsoft's
• 1:09
  PyRIT
• 1:09
  executes
• 1:09
  the
• 1:10
  campaigns,
• 1:10
  and
• 1:11
  results
• 1:11
  feed
• 1:11
  back
• 1:11
  for
• 1:11
  iterative
• 1:12
  refinement.
• 1:13
  In
• 1:13
  one
• 1:14
  engagement,
• 1:14
  the
• 1:15
  code-aware
• 1:15
  pipeline
• 1:15
  discovered
• 1:16
  that
• 1:16
  a
• 1:16
  guardrail
• 1:16
  confidence
• 1:17
  threshold
• 1:17
  was
• 1:18
  set
• 1:18
  to
• 1:18
  0.65,
• 1:19
  with
• 1:19
  attacks
• 1:20
  between
• 1:20
  0.40
• 1:21
  and
• 1:21
  0.65
• 1:23
  slipping
• 1:23
  through
• 1:23
  undetected
• 1:23
  —
• 1:24
  the
• 1:25
  kind
• 1:25
  of
• 1:25
  architecture-level
• 1:26
  finding
• 1:26
  that
• 1:26
  surface-level
• 1:27
  jailbreak
• 1:27
  probes
• 1:28
  never
• 1:28
  produce.
• 1:29
  Kutal
• 1:29
  notes
• 1:30
  the
• 1:30
  scanning
• 1:30
  market
• 1:30
  is
• 1:31
  consolidating
• 1:31
  fast
• 1:32
  (Promptfoo,
• 1:32
  Lakera,
• 1:33
  and
• 1:33
  SPLX
• 1:34
  all
• 1:34
  acquired),
• 1:35
  and
• 1:35
  argues
• 1:35
  the
• 1:36
  scanning
• 1:36
  layer
• 1:36
  is
• 1:36
  becoming
• 1:37
  a
• 1:37
  built-in
• 1:37
  feature
• 1:37
  while
• 1:38
  the
• 1:38
  expertise
• 1:38
  layer
• 1:38
  —
• 1:39
  understanding
• 1:40
  agent
• 1:40
  architectures,
• 1:41
  interpreting
• 1:42
  results
• 1:42
  in
• 1:42
  regulatory
• 1:43
  context
• 1:43
  —
• 1:44
  remains
• 1:44
  where
• 1:44
  judgment
• 1:45
  lives.
• 1:45
  Meanwhile,
• 1:46
  the
• 1:46
  nah
• 1:47
  project
• 1:47
  offers
• 1:47
  a
• 1:48
  pragmatic
• 1:48
  take
• 1:48
  on
• 1:48
  agent
• 1:49
  safety
• 1:49
  for
• 1:49
  Claude
• 1:49
  Code
• 1:50
  users.
• 1:51
  It
• 1:51
  intercepts
• 1:51
  every
• 1:52
  tool
• 1:52
  call
• 1:52
  through
• 1:52
  a
• 1:52
  PreToolUse
• 1:53
  hook
• 1:54
  and
• 1:54
  classifies
• 1:54
  it
• 1:54
  by
• 1:54
  what
• 1:55
  it
• 1:55
  actually
• 1:55
  does
• 1:55
  —
• 1:56
  structural
• 1:57
  command
• 1:57
  analysis,
• 1:58
  pipe
• 1:58
  composition,
• 1:59
  shell
• 1:59
  unwrapping
• 2:00
  —
• 2:01
  using
• 2:01
  deterministic
• 2:02
  rules
• 2:02
  that
• 2:02
  run
• 2:02
  in
• 2:03
  milliseconds
• 2:03
  before
• 2:04
  optionally
• 2:04
  routing
• 2:04
  ambiguous
• 2:05
  cases
• 2:05
  to
• 2:06
  an
• 2:06
  LLM.
• 2:07
  git
• 2:07
  push
• 2:07
  goes
• 2:08
  through;
• 2:08
  git
• 2:08
  push
• 2:09
  --force
• 2:09
  gets
• 2:10
  flagged.
• 2:10
  rm
• 2:11
  -rf
• 2:11
  pycache
• 2:12
  is
• 2:12
  fine;
• 2:12
  rm
• 2:13
  ~/.bashrc
• 2:14
  is
• 2:14
  blocked.
• 2:15
  The
• 2:15
  system
• 2:16
  self-protects:
• 2:17
  edits
• 2:17
  targeting
• 2:17
  the
• 2:18
  ~/.claude/hooks/
• 2:19
  directory
• 2:20
  are
• 2:20
  blocked
• 2:20
  outright,
• 2:21
  preventing
• 2:22
  the
• 2:22
  agent
• 2:22
  from
• 2:22
  disabling
• 2:23
  its
• 2:23
  own
• 2:23
  guardrails.
• 2:24
  Crucially,
• 2:25
  project-level
• 2:26
  .nah.yaml
• 2:28
  files
• 2:28
  can
• 2:28
  only
• 2:29
  tighten
• 2:29
  policies,
• 2:30
  never
• 2:30
  relax
• 2:30
  them
• 2:30
  —
• 2:31
  a
• 2:31
  supply-chain
• 2:32
  safety
• 2:32
  property
• 2:33
  that
• 2:33
  means
• 2:33
  a
• 2:33
  malicious
• 2:33
  repo
• 2:34
  cannot
• 2:34
  use
• 2:34
  configuration
• 2:35
  to
• 2:35
  allowlist
• 2:36
  dangerous
• 2:36
  commands.
• 2:37
  An
• 2:37
  interesting
• 2:38
  technical
• 2:38
  piece
• 2:38
  this
• 2:39
  week
• 2:39
  is
• 2:39
  the
• 2:39
  anatomy
• 2:40
  of
• 2:40
  Unicode
• 2:40
  ignorables
• 2:41
  —
• 2:41
  the
• 2:41
  4,174
• 2:43
  default-ignorable
• 2:44
  code
• 2:44
  points
• 2:44
  in
• 2:45
  Unicode
• 2:45
  15.1
• 2:46
  that
• 2:46
  render
• 2:46
  as
• 2:47
  nothing
• 2:47
  in
• 2:47
  most
• 2:47
  contexts
• 2:48
  but
• 2:48
  survive
• 2:49
  copy-paste,
• 2:50
  JSON
• 2:50
  serialization,
• 2:51
  database
• 2:51
  storage,
• 2:52
  and
• 2:52
  HTML
• 2:53
  transit
• 2:53
  to
• 2:53
  arrive
• 2:54
  intact
• 2:54
  at
• 2:54
  an
• 2:55
  LLM
• 2:55
  tokenizer.
• 2:56
  The
• 2:56
  tag
• 2:57
  block
• 2:57
  (U+E0000
• 2:59
  through
• 3:00
  U+E007F)
• 3:02
  alone
• 3:02
  provides
• 3:03
  128
• 3:04
  code
• 3:04
  points
• 3:04
  that
• 3:04
  can
• 3:04
  each
• 3:05
  hide
• 3:05
  one
• 3:05
  ASCII
• 3:05
  character,
• 3:06
  enough
• 3:07
  to
• 3:07
  embed
• 3:07
  entire
• 3:07
  instruction
• 3:08
  payloads
• 3:08
  invisible
• 3:09
  to
• 3:09
  human
• 3:09
  reviewers.
• 3:11
  The
• 3:11
  attack
• 3:11
  surface
• 3:11
  exists
• 3:12
  because
• 3:12
  a
• 3:12
  font
• 3:12
  renderer
• 3:13
  skips
• 3:13
  these
• 3:13
  characters
• 3:14
  while
• 3:14
  a
• 3:14
  BPE
• 3:14
  tokenizer
• 3:15
  processes
• 3:16
  the
• 3:16
  full
• 3:16
  byte
• 3:16
  sequence
• 3:16
  —
• 3:17
  the
• 3:17
  model
• 3:18
  sees
• 3:18
  instructions
• 3:19
  the
• 3:19
  user
• 3:19
  never
• 3:19
  saw.
• 3:21
  Variation
• 3:21
  selectors
• 3:22
  can
• 3:22
  shift
• 3:22
  tokenization
• 3:23
  boundaries,
• 3:24
  creating
• 3:24
  visually
• 3:25
  identical
• 3:25
  strings
• 3:25
  with
• 3:26
  different
• 3:26
  embeddings.
• 3:27
  Bidi
• 3:27
  controls
• 3:28
  can
• 3:28
  flip
• 3:28
  text
• 3:29
  direction
• 3:29
  to
• 3:29
  hide
• 3:29
  instruction
• 3:30
  fragments.
• 3:31
  The
• 3:31
  post
• 3:31
  catalogs
• 3:32
  concrete
• 3:32
  attack
• 3:33
  patterns:
• 3:34
  prompt
• 3:34
  injection
• 3:34
  via
• 3:35
  tag
• 3:35
  encoding,
• 3:36
  data
• 3:36
  exfiltration
• 3:37
  via
• 3:37
  invisible
• 3:38
  channels,
• 3:38
  cache
• 3:39
  poisoning
• 3:39
  through
• 3:39
  ZWJ-differentiated
• 3:41
  usernames.
• 3:42
  Defense
• 3:42
  requires
• 3:42
  actual
• 3:43
  Unicode
• 3:43
  property
• 3:44
  table
• 3:44
  lookups,
• 3:44
  not
• 3:45
  regex,
• 3:45
  and
• 3:46
  must
• 3:46
  check
• 3:46
  exact
• 3:46
  sequence
• 3:47
  membership
• 3:47
  before
• 3:48
  stripping
• 3:48
  —
• 3:48
  because
• 3:49
  some
• 3:49
  invisible
• 3:49
  characters
• 3:50
  are
• 3:50
  load-bearing
• 3:51
  (ZWJ
• 3:51
  in
• 3:52
  emoji
• 3:52
  sequences,
• 3:53
  variation
• 3:53
  selectors
• 3:54
  for
• 3:54
  presentation).
• 3:57
  Anthropic
• 3:57
  has
• 3:58
  moved
• 3:58
  the
• 3:58
  1M
• 3:58
  context
• 3:59
  window
• 3:59
  from
• 3:59
  premium
• 4:00
  feature
• 4:00
  to
• 4:00
  standard
• 4:01
  pricing
• 4:01
  across
• 4:01
  Claude
• 4:02
  Opus
• 4:02
  4.6
• 4:03
  and
• 4:03
  Sonnet
• 4:03
  4.6.
• 4:04
  No
• 4:05
  multiplier:
• 4:05
  a
• 4:05
  900K-token
• 4:06
  request
• 4:07
  is
• 4:07
  billed
• 4:07
  at
• 4:07
  the
• 4:07
  same
• 4:08
  per-token
• 4:08
  rate
• 4:08
  as
• 4:08
  a
• 4:08
  9K
• 4:09
  one,
• 4:09
  with
• 4:10
  full
• 4:10
  rate
• 4:10
  limits
• 4:10
  across
• 4:11
  the
• 4:11
  entire
• 4:11
  window.
• 4:12
  Media
• 4:12
  limits
• 4:13
  jump
• 4:13
  from
• 4:13
  100
• 4:13
  to
• 4:13
  600
• 4:14
  images
• 4:14
  or
• 4:14
  PDF
• 4:15
  pages
• 4:15
  per
• 4:15
  request.
• 4:16
  Claude
• 4:16
  Code
• 4:17
  users
• 4:17
  on
• 4:17
  Max,
• 4:18
  Team,
• 4:18
  and
• 4:18
  Enterprise
• 4:19
  plans
• 4:19
  with
• 4:19
  Opus
• 4:20
  4.6
• 4:21
  get
• 4:21
  the
• 4:21
  full
• 4:21
  window
• 4:21
  automatically,
• 4:22
  which
• 4:22
  means
• 4:23
  fewer
• 4:23
  compactions
• 4:24
  and
• 4:24
  more
• 4:24
  of
• 4:24
  the
• 4:24
  conversation
• 4:25
  kept
• 4:25
  intact.
• 4:26
  Opus
• 4:26
  4.6
• 4:27
  scores
• 4:27
  78.3%
• 4:28
  on
• 4:29
  MRCR
• 4:29
  v2
• 4:30
  at
• 4:30
  that
• 4:30
  context
• 4:30
  length,
• 4:31
  the
• 4:31
  highest
• 4:31
  among
• 4:32
  frontier
• 4:32
  models.
• 4:33
  The
• 4:33
  practical
• 4:34
  implications
• 4:34
  are
• 4:34
  immediate:
• 4:35
  Devin
• 4:36
  reports
• 4:36
  that
• 4:36
  large
• 4:37
  diffs
• 4:37
  that
• 4:37
  previously
• 4:37
  required
• 4:38
  chunking
• 4:38
  now
• 4:38
  fit
• 4:39
  in
• 4:39
  a
• 4:39
  single
• 4:39
  pass,
• 4:40
  producing
• 4:40
  higher-quality
• 4:41
  reviews
• 4:41
  from
• 4:41
  a
• 4:42
  simpler
• 4:42
  harness.
• 4:43
  One
• 4:43
  user
• 4:43
  reports
• 4:44
  a
• 4:44
  15%
• 4:45
  decrease
• 4:45
  in
• 4:45
  compaction
• 4:46
  events,
• 4:46
  meaning
• 4:47
  agents
• 4:47
  can
• 4:47
  run
• 4:47
  for
• 4:47
  hours
• 4:48
  without
• 4:48
  forgetting
• 4:49
  what
• 4:49
  they
• 4:49
  read
• 4:49
  on
• 4:49
  page
• 4:49
  one.
• 4:51
  NVIDIA's
• 4:51
  Nemotron
• 4:52
  3
• 4:52
  Super
• 4:52
  enters
• 4:53
  the
• 4:53
  same
• 4:53
  arena
• 4:53
  from
• 4:53
  the
• 4:54
  open-weight
• 4:54
  side
• 4:54
  —
• 4:55
  a
• 4:55
  120B
• 4:56
  mixture-of-experts
• 4:57
  hybrid
• 4:57
  with
• 4:58
  12B
• 4:58
  active
• 4:58
  parameters
• 4:59
  and
• 5:00
  a
• 5:00
  1M-token
• 5:00
  context
• 5:01
  window
• 5:01
  aimed
• 5:01
  at
• 5:02
  agentic
• 5:02
  workflows.
• 5:03
  The
• 5:03
  architecture
• 5:04
  choice
• 5:04
  is
• 5:04
  telling:
• 5:05
  MoE's
• 5:06
  sparse
• 5:06
  activation
• 5:07
  keeps
• 5:07
  per-token
• 5:07
  compute
• 5:08
  low
• 5:08
  enough
• 5:08
  to
• 5:08
  make
• 5:08
  million-token
• 5:09
  sessions
• 5:09
  economically
• 5:10
  viable,
• 5:11
  a
• 5:11
  constraint
• 5:12
  that
• 5:12
  matters
• 5:12
  enormously
• 5:13
  when
• 5:13
  an
• 5:13
  agent
• 5:14
  might
• 5:14
  chain
• 5:14
  hundreds
• 5:14
  of
• 5:14
  tool
• 5:15
  calls
• 5:15
  in
• 5:15
  a
• 5:15
  single
• 5:16
  run.
• 5:16
  On
• 5:17
  the
• 5:17
  training
• 5:17
  side,
• 5:17
  a
• 5:17
  new
• 5:18
  expert
• 5:18
  parallelism
• 5:19
  approach
• 5:19
  for
• 5:19
  fine-tuning
• 5:20
  trillion-parameter
• 5:21
  MoE
• 5:21
  models
• 5:21
  on
• 5:21
  a
• 5:21
  single
• 5:22
  8×H200
• 5:23
  node
• 5:23
  claims
• 5:23
  50×
• 5:23
  speed
• 5:24
  improvement
• 5:24
  and
• 5:24
  2×
• 5:25
  cost
• 5:25
  reduction
• 5:25
  over
• 5:25
  alternatives.
• 5:27
  The
• 5:27
  optimization
• 5:27
  journey
• 5:28
  from
• 5:28
  17
• 5:28
  seconds
• 5:29
  per
• 5:29
  step
• 5:29
  to
• 5:29
  2.86
• 5:30
  seconds
• 5:30
  covers
• 5:31
  grouped
• 5:31
  matmul,
• 5:32
  vectorized
• 5:32
  MXFP4
• 5:34
  dequantization,
• 5:35
  padding-aware
• 5:35
  token
• 5:36
  skipping,
• 5:36
  and
• 5:37
  sequence
• 5:37
  packing
• 5:37
  —
• 5:38
  the
• 5:38
  kind
• 5:38
  of
• 5:38
  systems-level
• 5:39
  engineering
• 5:40
  that
• 5:40
  turns
• 5:40
  "theoretically
• 5:41
  trainable"
• 5:41
  into
• 5:42
  "practically
• 5:42
  affordable."
• 5:43
  The
• 5:43
  authors
• 5:44
  note
• 5:44
  that
• 5:44
  gradient
• 5:45
  accumulation
• 5:45
  for
• 5:45
  MoE
• 5:46
  is
• 5:46
  still
• 5:46
  incorrectly
• 5:47
  implemented
• 5:47
  in
• 5:48
  several
• 5:48
  popular
• 5:48
  training
• 5:49
  frameworks,
• 5:50
  causing
• 5:50
  silent
• 5:50
  discrepancies
• 5:51
  between
• 5:52
  accumulated
• 5:52
  and
• 5:52
  equivalent
• 5:53
  batch
• 5:53
  sizes.
• 5:54
  Open-sourcing
• 5:55
  is
• 5:55
  promised
• 5:55
  in
• 5:55
  two
• 5:56
  weeks.
• 5:59
  The
• 5:59
  infrastructure
• 6:00
  layer
• 6:00
  beneath
• 6:00
  AI
• 6:01
  agents
• 6:01
  is
• 6:01
  filling
• 6:01
  in
• 6:01
  fast.
• 6:03
  GoClaw
• 6:03
  is
• 6:03
  a
• 6:03
  self-hosted
• 6:04
  agent
• 6:04
  gateway
• 6:04
  rewritten
• 6:05
  in
• 6:05
  Go
• 6:05
  as
• 6:05
  a
• 6:06
  single
• 6:06
  static
• 6:06
  binary
• 6:06
  —
• 6:07
  no
• 6:07
  Node.js,
• 6:08
  no
• 6:08
  npm,
• 6:08
  sub-50MB
• 6:09
  memory,
• 6:10
  under
• 6:10
  100ms
• 6:11
  cold
• 6:11
  start.
• 6:12
  It
• 6:12
  connects
• 6:12
  messaging
• 6:12
  channels
• 6:13
  (Telegram,
• 6:13
  WhatsApp,
• 6:14
  CLI)
• 6:14
  to
• 6:14
  any
• 6:14
  LLM
• 6:15
  provider
• 6:15
  (Claude,
• 6:16
  GPT,
• 6:16
  Gemini,
• 6:17
  DeepSeek,
• 6:17
  Ollama)
• 6:18
  and
• 6:18
  adds
• 6:18
  the
• 6:18
  plumbing
• 6:18
  that
• 6:18
  production
• 6:19
  deployments
• 6:19
  demand:
• 6:20
  multi-agent
• 6:20
  routing
• 6:21
  with
• 6:21
  declarative
• 6:21
  bindings,
• 6:22
  inter-agent
• 6:22
  delegation
• 6:23
  via
• 6:23
  an
• 6:23
  ask
• 6:24
  agent
• 6:24
  tool,
• 6:24
  persistent
• 6:25
  memory
• 6:25
  with
• 6:25
  BM25
• 6:26
  search,
• 6:27
  a
• 6:27
  skill
• 6:27
  system
• 6:27
  using
• 6:27
  Markdown
• 6:28
  files
• 6:28
  with
• 6:28
  YAML
• 6:29
  frontmatter,
• 6:29
  heartbeat
• 6:30
  daemons
• 6:30
  for
• 6:30
  proactive
• 6:31
  agent
• 6:31
  actions,
• 6:31
  and
• 6:32
  cron
• 6:32
  scheduling
• 6:32
  with
• 6:32
  runtime
• 6:33
  management.
• 6:34
  Security
• 6:34
  is
• 6:34
  thoughtfully
• 6:35
  layered
• 6:35
  —
• 6:36
  localhost-only
• 6:37
  binding
• 6:37
  by
• 6:37
  default,
• 6:38
  bearer
• 6:38
  token
• 6:38
  auth
• 6:38
  with
• 6:39
  constant-time
• 6:39
  comparison,
• 6:40
  SSRF
• 6:41
  protection
• 6:41
  that
• 6:41
  re-validates
• 6:42
  redirect
• 6:42
  targets
• 6:42
  at
• 6:42
  each
• 6:43
  hop,
• 6:43
  workspace
• 6:44
  containment
• 6:44
  with
• 6:44
  symlink
• 6:44
  resolution,
• 6:45
  and
• 6:46
  exec
• 6:46
  approval
• 6:46
  policies
• 6:47
  that
• 6:47
  block
• 6:47
  shell
• 6:47
  metacharacters
• 6:48
  in
• 6:48
  allowlist
• 6:48
  mode.
• 6:49
  The
• 6:49
  feature
• 6:50
  comparison
• 6:50
  with
• 6:50
  its
• 6:50
  Node.js
• 6:52
  predecessor
• 6:52
  OpenClaw
• 6:53
  shows
• 6:53
  parity
• 6:53
  on
• 6:53
  most
• 6:54
  capabilities
• 6:54
  while
• 6:54
  adding
• 6:55
  inter-agent
• 6:55
  delegation
• 6:56
  and
• 6:56
  eliminating
• 6:57
  the
• 6:57
  runtime
• 6:57
  dependency
• 6:58
  chain.
• 6:59
  Axon
• 6:59
  approaches
• 6:59
  agent
• 7:00
  intelligence
• 7:00
  from
• 7:00
  the
• 7:00
  code-understanding
• 7:01
  side.
• 7:02
  It
• 7:02
  indexes
• 7:02
  any
• 7:03
  Python,
• 7:03
  TypeScript,
• 7:04
  or
• 7:04
  JavaScript
• 7:04
  codebase
• 7:05
  into
• 7:05
  a
• 7:05
  structural
• 7:05
  knowledge
• 7:06
  graph
• 7:06
  —
• 7:07
  12
• 7:07
  analysis
• 7:08
  phases
• 7:08
  covering
• 7:08
  parsing,
• 7:09
  call
• 7:09
  tracing,
• 7:09
  type
• 7:10
  analysis,
• 7:10
  community
• 7:11
  detection
• 7:11
  via
• 7:11
  the
• 7:12
  Leiden
• 7:12
  algorithm,
• 7:13
  execution
• 7:13
  flow
• 7:13
  tracing,
• 7:14
  dead
• 7:14
  code
• 7:14
  detection,
• 7:15
  git
• 7:15
  change
• 7:15
  coupling,
• 7:16
  and
• 7:16
  384-dimensional
• 7:17
  embeddings.
• 7:19
  The
• 7:19
  result
• 7:19
  is
• 7:20
  an
• 7:20
  MCP
• 7:20
  server
• 7:21
  where
• 7:21
  axon
• 7:21
  impact("validate")
• 7:22
  returns
• 7:23
  all
• 7:23
  47
• 7:24
  affected
• 7:24
  symbols
• 7:24
  grouped
• 7:25
  by
• 7:25
  depth
• 7:25
  (will
• 7:25
  break
• 7:25
  /
• 7:25
  may
• 7:26
  break
• 7:26
  /
• 7:26
  review)
• 7:26
  with
• 7:26
  confidence
• 7:27
  scores,
• 7:27
  in
• 7:27
  a
• 7:27
  single
• 7:28
  tool
• 7:28
  call.
• 7:29
  That
• 7:29
  is
• 7:30
  qualitatively
• 7:30
  different
• 7:31
  from
• 7:31
  an
• 7:31
  agent
• 7:31
  grepping
• 7:31
  for
• 7:31
  callers
• 7:32
  and
• 7:32
  hoping
• 7:32
  it
• 7:32
  found
• 7:33
  them
• 7:33
  all.
• 7:34
  The
• 7:34
  web
• 7:34
  dashboard
• 7:34
  adds
• 7:35
  force-directed
• 7:35
  graph
• 7:36
  visualization
• 7:36
  with
• 7:36
  Sigma.js,
• 7:37
  a
• 7:38
  Cypher
• 7:38
  console,
• 7:39
  and
• 7:39
  coupling
• 7:39
  heatmaps
• 7:39
  —
• 7:40
  useful
• 7:41
  for
• 7:41
  humans
• 7:41
  and
• 7:41
  agents
• 7:42
  alike.
• 7:43
  Christian
• 7:43
  Posta's
• 7:43
  AAuth
• 7:44
  demo
• 7:44
  walks
• 7:44
  through
• 7:44
  a
• 7:44
  concrete
• 7:45
  implementation
• 7:45
  of
• 7:45
  agent
• 7:46
  identity
• 7:46
  and
• 7:46
  access
• 7:47
  management
• 7:47
  using
• 7:47
  Keycloak
• 7:48
  and
• 7:48
  Agentgateway,
• 7:49
  with
• 7:49
  libraries
• 7:49
  in
• 7:49
  Java,
• 7:50
  Python,
• 7:50
  and
• 7:50
  Rust.
• 7:52
  The
• 7:52
  flows
• 7:52
  cover
• 7:52
  pseudonymous
• 7:53
  access
• 7:53
  via
• 7:53
  Header
• 7:54
  Web
• 7:54
  Keys,
• 7:55
  identified
• 7:55
  access
• 7:56
  via
• 7:56
  JWKS,
• 7:57
  and
• 7:57
  user
• 7:57
  consent
• 7:58
  with
• 7:58
  authorization
• 7:58
  —
• 7:59
  the
• 7:59
  kinds
• 8:00
  of
• 8:00
  boring,
• 8:01
  essential
• 8:01
  plumbing
• 8:01
  that
• 8:01
  prevents
• 8:02
  the
• 8:02
  agent
• 8:02
  ecosystem
• 8:03
  from
• 8:03
  becoming
• 8:03
  a
• 8:03
  permissioning
• 8:04
  free-for-all.
• 8:05
  For
• 8:05
  observability,
• 8:06
  an
• 8:06
  eBPF-based
• 8:07
  tracer
• 8:08
  called
• 8:08
  agtap
• 8:08
  captures
• 8:09
  LLM
• 8:09
  API
• 8:10
  and
• 8:10
  MCP
• 8:10
  traffic
• 8:11
  from
• 8:11
  any
• 8:11
  process
• 8:12
  on
• 8:12
  a
• 8:12
  Linux
• 8:12
  machine
• 8:12
  with
• 8:12
  zero
• 8:13
  code
• 8:13
  changes
• 8:13
  —
• 8:14
  no
• 8:14
  SDK
• 8:15
  modifications,
• 8:16
  no
• 8:16
  proxy.
• 8:17
  It
• 8:17
  intercepts
• 8:18
  TLS
• 8:18
  via
• 8:19
  OpenSSL
• 8:19
  uprobes
• 8:20
  and
• 8:20
  parses
• 8:20
  Anthropic,
• 8:21
  OpenAI,
• 8:22
  and
• 8:22
  Gemini
• 8:22
  API
• 8:23
  calls
• 8:23
  in
• 8:23
  real
• 8:23
  time,
• 8:24
  extracting
• 8:25
  model
• 8:25
  names,
• 8:26
  token
• 8:26
  counts,
• 8:26
  latency,
• 8:27
  time-to-first-token,
• 8:28
  tool
• 8:28
  names,
• 8:29
  and
• 8:29
  streaming
• 8:29
  status.
• 8:31
  Output
• 8:31
  goes
• 8:31
  to
• 8:31
  JSONL,
• 8:32
  OpenTelemetry,
• 8:33
  or
• 8:33
  Prometheus.
• 8:35
  The
• 8:35
  limitation:
• 8:36
  it
• 8:36
  requires
• 8:37
  root
• 8:37
  and
• 8:37
  only
• 8:37
  works
• 8:38
  with
• 8:38
  OpenSSL-linked
• 8:39
  processes,
• 8:40
  excluding
• 8:40
  Go,
• 8:41
  Bun,
• 8:41
  Deno,
• 8:41
  and
• 8:41
  rustls.
• 8:43
  A
• 8:43
  short
• 8:43
  video
• 8:43
  from
• 8:43
  Chase
• 8:44
  argues
• 8:44
  for
• 8:44
  replacing
• 8:45
  MCP
• 8:45
  servers
• 8:46
  with
• 8:46
  CLI
• 8:46
  tools
• 8:47
  inside
• 8:47
  Claude
• 8:47
  Code,
• 8:48
  pointing
• 8:49
  to
• 8:49
  the
• 8:49
  CLI-Anything
• 8:50
  open-source
• 8:50
  project
• 8:51
  from
• 8:51
  the
• 8:51
  LightRAG
• 8:51
  team
• 8:52
  that
• 8:52
  auto-generates
• 8:53
  CLI
• 8:53
  wrappers
• 8:54
  from
• 8:54
  any
• 8:54
  open-source
• 8:55
  project.
• 8:56
  The
• 8:56
  thesis:
• 8:57
  CLIs
• 8:58
  are
• 8:58
  lighter
• 8:58
  on
• 8:58
  tokens
• 8:59
  and
• 8:59
  context
• 8:59
  than
• 8:59
  MCP's
• 9:00
  JSON-RPC
• 9:03
  overhead.
• 9:04
  It
• 9:04
  is
• 9:04
  a
• 9:04
  valid
• 9:05
  optimization
• 9:05
  for
• 9:05
  some
• 9:06
  workflows,
• 9:07
  though
• 9:07
  it
• 9:07
  trades
• 9:07
  away
• 9:07
  the
• 9:07
  structured
• 9:08
  tool
• 9:08
  discovery
• 9:09
  and
• 9:09
  streaming
• 9:09
  capabilities
• 9:10
  that
• 9:10
  make
• 9:10
  MCP
• 9:11
  useful
• 9:11
  for
• 9:11
  complex
• 9:12
  agent
• 9:12
  orchestration.
• 9:15
  A
• 9:15
  threat
• 9:16
  actor
• 9:16
  calling
• 9:16
  themselves
• 9:17
  ByteToBreach
• 9:17
  has
• 9:18
  leaked
• 9:18
  what
• 9:18
  they
• 9:18
  claim
• 9:19
  is
• 9:19
  the
• 9:19
  complete
• 9:19
  source
• 9:19
  code
• 9:20
  of
• 9:20
  Sweden's
• 9:20
  e-government
• 9:21
  platform,
• 9:21
  obtained
• 9:22
  through
• 9:22
  compromised
• 9:22
  CGI
• 9:23
  Sverige
• 9:23
  AB
• 9:24
  infrastructure.
• 9:25
  CGI
• 9:25
  Sverige
• 9:26
  is
• 9:26
  the
• 9:26
  Swedish
• 9:26
  subsidiary
• 9:27
  of
• 9:27
  global
• 9:27
  IT
• 9:28
  services
• 9:28
  giant
• 9:29
  CGI
• 9:29
  Group
• 9:30
  and
• 9:30
  manages
• 9:30
  critical
• 9:31
  government
• 9:31
  digital
• 9:31
  services.
• 9:32
  The
• 9:32
  disclosed
• 9:33
  attack
• 9:33
  chain
• 9:33
  includes
• 9:34
  a
• 9:34
  full
• 9:34
  Jenkins
• 9:34
  compromise,
• 9:35
  Docker
• 9:35
  escape
• 9:36
  via
• 9:36
  the
• 9:36
  Jenkins
• 9:36
  user
• 9:37
  being
• 9:37
  in
• 9:37
  the
• 9:37
  Docker
• 9:37
  group,
• 9:38
  SSH
• 9:38
  private
• 9:39
  key
• 9:39
  pivots,
• 9:40
  and
• 9:40
  SQL
• 9:40
  copy-to-program
• 9:41
  exploitation.
• 9:42
  The
• 9:42
  actor
• 9:43
  is
• 9:43
  releasing
• 9:43
  the
• 9:43
  source
• 9:44
  code
• 9:44
  for
• 9:44
  free
• 9:44
  while
• 9:45
  selling
• 9:45
  citizen
• 9:45
  PII
• 9:46
  databases
• 9:47
  and
• 9:47
  electronic
• 9:47
  signing
• 9:48
  documents
• 9:48
  separately
• 9:48
  —
• 9:50
  a
• 9:50
  deliberate
• 9:50
  strategy
• 9:51
  to
• 9:51
  maximize
• 9:51
  exposure
• 9:52
  while
• 9:52
  monetizing
• 9:53
  the
• 9:53
  most
• 9:53
  sensitive
• 9:53
  data.
• 9:54
  The
• 9:55
  same
• 9:55
  actor
• 9:55
  was
• 9:55
  behind
• 9:56
  the
• 9:56
  Viking
• 9:56
  Line
• 9:56
  breach
• 9:57
  posted
• 9:57
  the
• 9:57
  day
• 9:57
  before,
• 9:58
  and
• 9:58
  explicitly
• 9:59
  called
• 9:59
  out
• 9:59
  companies
• 10:00
  that
• 10:00
  blame
• 10:00
  breaches
• 10:00
  on
• 10:00
  third
• 10:01
  parties,
• 10:01
  stating
• 10:02
  this
• 10:02
  compromise
• 10:03
  "belongs
• 10:03
  clearly
• 10:03
  to
• 10:03
  CGI
• 10:04
  infrastructure."
• 10:05
  For
• 10:05
  anyone
• 10:06
  in
• 10:06
  government
• 10:06
  IT
• 10:07
  contracting,
• 10:08
  the
• 10:08
  Jenkins-to-Docker-escape-to-SSH-pivot
• 10:09
  chain
• 10:10
  is
• 10:10
  a
• 10:10
  reminder
• 10:10
  that
• 10:10
  container
• 10:11
  boundaries
• 10:11
  are
• 10:11
  only
• 10:11
  as
• 10:12
  strong
• 10:12
  as
• 10:12
  the
• 10:12
  user
• 10:12
  permissions
• 10:13
  that
• 10:13
  cross
• 10:13
  them.
• 10:14
  After
• 10:14
  nearly
• 10:14
  a
• 10:14
  decade
• 10:15
  of
• 10:15
  work
• 10:15
  with
• 10:15
  AWS
• 10:16
  security
• 10:16
  teams,
• 10:17
  the
• 10:17
  bucketsquatting
• 10:17
  problem
• 10:18
  in
• 10:18
  S3
• 10:18
  finally
• 10:19
  has
• 10:19
  a
• 10:19
  structural
• 10:19
  fix.
• 10:20
  AWS
• 10:21
  introduced
• 10:21
  account-namespaced
• 10:22
  bucket
• 10:22
  names
• 10:23
  —
• 10:23
  ------x-s3
• 10:23
  —
• 10:24
  where
• 10:24
  only
• 10:25
  the
• 10:25
  owning
• 10:25
  account
• 10:25
  can
• 10:25
  create
• 10:26
  buckets
• 10:26
  in
• 10:26
  that
• 10:26
  namespace.
• 10:27
  The
• 10:27
  old
• 10:28
  problem:
• 10:28
  S3
• 10:29
  bucket
• 10:29
  names
• 10:29
  are
• 10:29
  globally
• 10:30
  unique,
• 10:30
  and
• 10:31
  if
• 10:31
  you
• 10:31
  delete
• 10:31
  one,
• 10:31
  anyone
• 10:32
  can
• 10:32
  register
• 10:32
  it.
• 10:33
  Predictable
• 10:34
  naming
• 10:34
  conventions
• 10:35
  (appending
• 10:35
  region
• 10:36
  names)
• 10:36
  made
• 10:36
  it
• 10:36
  trivial
• 10:36
  for
• 10:37
  attackers
• 10:37
  to
• 10:37
  squat
• 10:38
  on
• 10:38
  buckets
• 10:38
  that
• 10:38
  organizations
• 10:39
  or
• 10:39
  AWS's
• 10:40
  own
• 10:40
  internal
• 10:40
  teams
• 10:41
  might
• 10:41
  recreate.
• 10:42
  The
• 10:42
  new
• 10:42
  namespace
• 10:43
  is
• 10:43
  recommended
• 10:44
  by
• 10:44
  default,
• 10:44
  enforceable
• 10:45
  via
• 10:45
  an
• 10:45
  SCP
• 10:46
  condition
• 10:46
  key
• 10:46
  s3:ResourceAccount,
• 10:49
  and
• 10:49
  available
• 10:49
  on
• 10:49
  all
• 10:50
  major
• 10:50
  cloud
• 10:50
  providers.
• 10:51
  It
• 10:51
  does
• 10:51
  not
• 10:52
  retroactively
• 10:52
  protect
• 10:53
  existing
• 10:53
  buckets
• 10:53
  —
• 10:54
  migration
• 10:54
  is
• 10:55
  manual.
• 10:56
  Google
• 10:56
  Cloud
• 10:56
  Storage
• 10:57
  already
• 10:57
  scopes
• 10:57
  buckets
• 10:58
  by
• 10:58
  verified
• 10:59
  domain
• 10:59
  ownership.
• 11:00
  Azure
• 11:00
  Blob
• 11:01
  Storage
• 11:01
  remains
• 11:01
  vulnerable
• 11:02
  due
• 11:02
  to
• 11:02
  its
• 11:02
  24-character
• 11:03
  account
• 11:03
  name
• 11:04
  limit
• 11:04
  creating
• 11:05
  a
• 11:05
  constrained
• 11:05
  namespace.
• 11:06
  In
• 11:07
  a
• 11:07
  quieter
• 11:07
  but
• 11:07
  consequential
• 11:08
  move,
• 11:08
  Instagram
• 11:09
  will
• 11:09
  drop
• 11:09
  end-to-end
• 11:10
  encryption
• 11:10
  for
• 11:10
  messaging
• 11:11
  after
• 11:11
  May
• 11:11
  8,
• 11:12
  2026.
• 11:13
  The
• 11:13
  help
• 11:13
  page
• 11:13
  offers
• 11:14
  a
• 11:14
  single
• 11:14
  sentence
• 11:14
  of
• 11:15
  justification:
• 11:16
  E2E
• 11:16
  encryption
• 11:17
  "adds
• 11:17
  extra
• 11:17
  security
• 11:18
  and
• 11:18
  protection."
• 11:19
  Removing
• 11:19
  it
• 11:19
  subtracts
• 11:20
  that
• 11:20
  protection.
• 11:21
  For
• 11:21
  a
• 11:21
  platform
• 11:22
  with
• 11:22
  over
• 11:22
  two
• 11:22
  billion
• 11:22
  users,
• 11:23
  the
• 11:23
  regression
• 11:24
  is
• 11:24
  notable
• 11:24
  —
• 11:25
  particularly
• 11:25
  given
• 11:26
  that
• 11:26
  Meta
• 11:26
  spent
• 11:26
  years
• 11:27
  implementing
• 11:27
  E2E
• 11:28
  across
• 11:28
  its
• 11:28
  messaging
• 11:28
  products
• 11:29
  after
• 11:29
  sustained
• 11:30
  pressure
• 11:30
  from
• 11:30
  privacy
• 11:30
  advocates.
• 11:34
  Prateek
• 11:34
  Karnal's
• 11:34
  detailed
• 11:35
  account
• 11:35
  of
• 11:35
  building
• 11:35
  a
• 11:36
  decentralized
• 11:36
  self-improving
• 11:37
  AI
• 11:37
  system
• 11:38
  offers
• 11:38
  the
• 11:38
  most
• 11:39
  concrete
• 11:39
  operational
• 11:40
  data
• 11:40
  we
• 11:40
  have
• 11:40
  seen
• 11:40
  on
• 11:40
  agentic
• 11:41
  orchestration
• 11:41
  at
• 11:41
  scale.
• 11:43
  The
• 11:43
  setup:
• 11:43
  35
• 11:44
  parallel
• 11:44
  agent
• 11:45
  sessions
• 11:45
  running
• 11:45
  overnight,
• 11:46
  orchestrated
• 11:47
  through
• 11:47
  OpenClaw
• 11:47
  on
• 11:47
  Telegram,
• 11:48
  executing
• 11:48
  on
• 11:49
  Agent
• 11:49
  Orchestrator,
• 11:49
  with
• 11:50
  a
• 11:50
  human
• 11:50
  as
• 11:50
  the
• 11:50
  final
• 11:51
  governance
• 11:51
  layer.
• 11:51
  Two-tier
• 11:52
  review
• 11:52
  —
• 11:53
  Codex
• 11:53
  for
• 11:53
  fast
• 11:54
  first-pass,
• 11:54
  Claude
• 11:54
  Code
• 11:55
  for
• 11:55
  deep
• 11:55
  architectural
• 11:56
  analysis
• 11:56
  —
• 11:57
  produced
• 11:57
  dependency-aware
• 11:57
  merge
• 11:58
  ordering
• 11:58
  across
• 11:58
  six
• 11:59
  PRs.
• 12:00
  The
• 12:00
  system
• 12:00
  dogfooded
• 12:00
  its
• 12:01
  own
• 12:01
  feedback
• 12:01
  tools:
• 12:02
  session
• 12:02
  ao-22
• 12:03
  filed
• 12:03
  bug
• 12:03
  reports
• 12:04
  using
• 12:04
  contracts
• 12:04
  it
• 12:04
  had
• 12:05
  designed,
• 12:05
  OpenClaw
• 12:05
  verified
• 12:06
  delivery,
• 12:07
  the
• 12:07
  human
• 12:07
  caught
• 12:07
  a
• 12:07
  severity
• 12:07
  classification
• 12:08
  error,
• 12:09
  ao-22
• 12:09
  fixed
• 12:10
  the
• 12:10
  classification
• 12:10
  logic,
• 12:11
  and
• 12:11
  ao-24
• 12:12
  re-validated.
• 12:13
  Four
• 12:13
  feedback
• 12:14
  layers,
• 12:14
  three
• 12:14
  different
• 12:14
  agents,
• 12:15
  one
• 12:15
  human
• 12:15
  judgment
• 12:16
  call,
• 12:16
  twelve
• 12:16
  minutes
• 12:16
  elapsed.
• 12:18
  The
• 12:18
  codebase
• 12:18
  grew
• 12:18
  from
• 12:18
  43K
• 12:19
  to
• 12:19
  82K
• 12:20
  lines
• 12:20
  in
• 12:20
  19
• 12:20
  days.
• 12:21
  But
• 12:21
  the
• 12:21
  failures
• 12:21
  are
• 12:21
  instructive:
• 12:22
  the
• 12:22
  dashboard
• 12:23
  crashed
• 12:23
  five
• 12:23
  times,
• 12:24
  GitHub's
• 12:24
  GraphQL
• 12:25
  API
• 12:25
  rate-limited
• 12:26
  PR
• 12:26
  creation,
• 12:27
  API
• 12:27
  token
• 12:27
  exhaustion
• 12:28
  caused
• 12:28
  stream
• 12:29
  disconnects
• 12:29
  across
• 12:29
  simultaneous
• 12:30
  sessions,
• 12:30
  and
• 12:31
  authentication
• 12:31
  failures
• 12:32
  blocked
• 12:32
  review
• 12:32
  sessions.
• 12:33
  Every
• 12:34
  failure
• 12:34
  became
• 12:34
  a
• 12:34
  structured
• 12:35
  report
• 12:35
  that
• 12:35
  fed
• 12:35
  back
• 12:36
  into
• 12:36
  the
• 12:36
  system
• 12:36
  —
• 12:37
  the
• 12:37
  dashboard
• 12:38
  crashes
• 12:38
  spawned
• 12:38
  ao-27,
• 12:39
  which
• 12:40
  shipped
• 12:40
  594
• 12:41
  tests
• 12:41
  for
• 12:42
  reliability.
• 12:43
  The
• 12:43
  governance
• 12:43
  model
• 12:44
  enforces
• 12:44
  consent
• 12:44
  gates:
• 12:45
  no
• 12:45
  agent
• 12:45
  can
• 12:46
  autonomously
• 12:46
  fork
• 12:46
  a
• 12:46
  project,
• 12:47
  open
• 12:47
  a
• 12:47
  PR,
• 12:48
  or
• 12:48
  change
• 12:48
  push
• 12:48
  targets
• 12:49
  without
• 12:49
  human
• 12:49
  approval.
• 12:50
  Volodymyr
• 12:51
  Dybenko
• 12:51
  describes
• 12:52
  a
• 12:52
  lighter-weight
• 12:52
  version
• 12:52
  of
• 12:52
  the
• 12:53
  same
• 12:53
  pattern:
• 12:53
  feeding
• 12:54
  Codex
• 12:54
  output
• 12:55
  to
• 12:55
  Claude
• 12:55
  for
• 12:55
  architectural
• 12:56
  critique,
• 12:56
  then
• 12:57
  feeding
• 12:57
  Claude's
• 12:57
  refined
• 12:58
  task
• 12:58
  back
• 12:58
  to
• 12:58
  a
• 12:58
  Gemini-based
• 12:59
  executor.
• 13:00
  In
• 13:00
  one
• 13:01
  evening,
• 13:01
  five
• 13:02
  of
• 13:02
  ten
• 13:02
  major
• 13:03
  architectural
• 13:03
  changes
• 13:03
  completed
• 13:04
  —
• 13:05
  work
• 13:05
  estimated
• 13:05
  at
• 13:05
  two
• 13:05
  months
• 13:06
  of
• 13:06
  engineering
• 13:06
  time.
• 13:07
  The
• 13:07
  key
• 13:08
  insight
• 13:08
  is
• 13:08
  not
• 13:08
  that
• 13:09
  AI
• 13:09
  writes
• 13:09
  code
• 13:09
  fast,
• 13:10
  but
• 13:10
  that
• 13:10
  AI
• 13:10
  systems
• 13:11
  reviewing,
• 13:11
  challenging,
• 13:12
  and
• 13:12
  improving
• 13:13
  each
• 13:13
  other's
• 13:13
  thinking
• 13:13
  creates
• 13:14
  a
• 13:14
  compound
• 13:14
  effect.
• 13:15
  The
• 13:15
  builder
• 13:16
  becomes
• 13:16
  the
• 13:16
  orchestrator.
• 13:17
  A
• 13:17
  detailed
• 13:18
  video
• 13:18
  walkthrough
• 13:19
  of
• 13:19
  a
• 13:19
  self-healing
• 13:19
  AI
• 13:20
  coding
• 13:20
  workflow
• 13:20
  packages
• 13:21
  the
• 13:21
  validation
• 13:21
  problem
• 13:22
  into
• 13:22
  a
• 13:22
  single
• 13:22
  Claude
• 13:23
  Code
• 13:23
  skill
• 13:23
  command.
• 13:24
  The
• 13:24
  approach
• 13:25
  delegates
• 13:25
  validation
• 13:26
  to
• 13:26
  the
• 13:26
  coding
• 13:26
  agent
• 13:27
  itself:
• 13:28
  sub-agents
• 13:28
  run
• 13:28
  parallel
• 13:29
  research
• 13:29
  mapping
• 13:30
  codebase
• 13:30
  structure
• 13:31
  and
• 13:31
  user
• 13:31
  journeys,
• 13:32
  then
• 13:32
  a
• 13:32
  browser
• 13:32
  automation
• 13:33
  tool
• 13:33
  (Vercel's
• 13:33
  agent
• 13:34
  browser
• 13:34
  CLI)
• 13:35
  systematically
• 13:35
  walks
• 13:36
  each
• 13:36
  user
• 13:36
  flow,
• 13:37
  taking
• 13:37
  snapshots
• 13:38
  and
• 13:38
  querying
• 13:38
  after
• 13:39
  write
• 13:39
  operations.
• 13:40
  The
• 13:40
  plan-implement-validate
• 13:41
  loop
• 13:42
  runs
• 13:42
  automatically
• 13:43
  on
• 13:43
  every
• 13:43
  implementation
• 13:44
  step.
• 13:45
  It
• 13:45
  is
• 13:45
  token-heavy
• 13:46
  but
• 13:46
  addresses
• 13:47
  the
• 13:47
  fundamental
• 13:47
  bottleneck:
• 13:48
  AI-generated
• 13:49
  code
• 13:49
  ships
• 13:50
  faster
• 13:50
  than
• 13:50
  humans
• 13:51
  can
• 13:51
  review
• 13:51
  it,
• 13:52
  and
• 13:52
  the
• 13:52
  agent
• 13:52
  that
• 13:52
  wrote
• 13:52
  the
• 13:52
  code
• 13:53
  is
• 13:53
  best
• 13:53
  positioned
• 13:54
  to
• 13:54
  verify
• 13:54
  it
• 13:54
  works
• 13:54
  —
• 13:55
  provided
• 13:56
  you
• 13:56
  give
• 13:56
  it
• 13:56
  a
• 13:56
  structured
• 13:57
  framework
• 13:57
  rather
• 13:58
  than
• 13:58
  trusting
• 13:58
  it
• 13:58
  to
• 13:58
  self-assess.
• 14:02
  Rob
• 14:03
  T.
• 14:03
  Lee's
• 14:03
  essay
• 14:03
  draws
• 14:04
  a
• 14:04
  sharp
• 14:04
  parallel
• 14:04
  between
• 14:05
  the
• 14:05
  current
• 14:05
  state
• 14:05
  of
• 14:05
  AI
• 14:06
  security
• 14:06
  and
• 14:06
  cybersecurity
• 14:07
  circa
• 14:07
  1997.
• 14:08
  SANS
• 14:09
  was
• 14:09
  founded
• 14:09
  in
• 14:09
  1989
• 14:10
  as
• 14:10
  a
• 14:10
  cooperative
• 14:11
  for
• 14:11
  sysadmins
• 14:11
  who
• 14:11
  also
• 14:12
  needed
• 14:12
  to
• 14:12
  protect
• 14:12
  their
• 14:12
  networks
• 14:13
  —
• 14:14
  security
• 14:14
  was
• 14:14
  an
• 14:14
  afterthought
• 14:15
  with
• 14:15
  a
• 14:15
  comma.
• 14:16
  The
• 14:16
  CISO
• 14:16
  title
• 14:17
  did
• 14:17
  not
• 14:17
  widely
• 14:17
  exist
• 14:18
  until
• 14:18
  the
• 14:18
  mid-2000s,
• 14:19
  a
• 14:19
  decade
• 14:20
  after
• 14:20
  the
• 14:20
  commercial
• 14:20
  internet.
• 14:21
  US
• 14:21
  Cyber
• 14:22
  Command
• 14:22
  stood
• 14:22
  up
• 14:22
  in
• 14:22
  2009.
• 14:24
  The
• 14:24
  pattern:
• 14:25
  a
• 14:25
  15-to-20-year
• 14:26
  lag
• 14:26
  between
• 14:26
  a
• 14:26
  technology
• 14:27
  becoming
• 14:27
  real
• 14:28
  and
• 14:28
  a
• 14:28
  workforce
• 14:29
  structure
• 14:29
  forming
• 14:29
  around
• 14:30
  it.
• 14:31
  AI
• 14:31
  security,
• 14:32
  Lee
• 14:32
  argues,
• 14:33
  does
• 14:33
  not
• 14:33
  get
• 14:33
  15
• 14:34
  years
• 14:34
  —
• 14:35
  nation-state
• 14:35
  actors
• 14:35
  were
• 14:36
  using
• 14:36
  AI
• 14:36
  autonomously
• 14:37
  for
• 14:37
  cyber
• 14:37
  operations
• 14:38
  before
• 14:38
  most
• 14:38
  organizations
• 14:39
  had
• 14:39
  an
• 14:39
  AI
• 14:40
  security
• 14:40
  policy.
• 14:41
  Whether
• 14:41
  AI
• 14:41
  security
• 14:42
  ends
• 14:42
  up
• 14:42
  as
• 14:42
  a
• 14:42
  subspecialty
• 14:43
  of
• 14:43
  traditional
• 14:44
  cybersecurity
• 14:44
  or
• 14:44
  a
• 14:44
  genuinely
• 14:45
  distinct
• 14:45
  field
• 14:46
  remains
• 14:46
  undefined.
• 14:47
  The
• 14:47
  most
• 14:47
  important
• 14:48
  job
• 14:48
  in
• 14:48
  AI
• 14:48
  security
• 14:49
  in
• 14:49
  2030
• 14:50
  may
• 14:50
  not
• 14:50
  have
• 14:50
  a
• 14:50
  name
• 14:50
  yet.
• 14:51
  Steve
• 14:51
  Yegge's
• 14:52
  wide-ranging
• 14:52
  conversation
• 14:53
  on
• 14:53
  AI
• 14:53
  adoption
• 14:54
  levels
• 14:54
  puts
• 14:54
  numbers
• 14:54
  to
• 14:55
  the
• 14:55
  workforce
• 14:55
  divide.
• 14:56
  His
• 14:56
  eight-level
• 14:56
  framework
• 14:57
  runs
• 14:57
  from
• 14:57
  no
• 14:58
  AI
• 14:58
  (level
• 14:58
  one)
• 14:58
  through
• 14:59
  IDE-integrated
• 14:59
  yes/no
• 15:00
  approvals
• 15:00
  (level
• 15:01
  two)
• 15:01
  up
• 15:01
  to
• 15:01
  running
• 15:01
  multiple
• 15:02
  agents
• 15:02
  in
• 15:02
  parallel
• 15:03
  while
• 15:03
  you
• 15:03
  nap
• 15:03
  (level
• 15:04
  six
• 15:04
  and
• 15:04
  above).
• 15:04
  His
• 15:05
  estimate:
• 15:05
  70%
• 15:06
  of
• 15:06
  engineers
• 15:07
  are
• 15:07
  stuck
• 15:07
  at
• 15:07
  the
• 15:07
  bottom
• 15:07
  levels.
• 15:09
  The
• 15:09
  "vampiric
• 15:09
  effect"
• 15:09
  —
• 15:10
  AI
• 15:11
  makes
• 15:11
  you
• 15:11
  intensely
• 15:11
  productive
• 15:12
  for
• 15:12
  short
• 15:12
  bursts,
• 15:13
  then
• 15:13
  you
• 15:13
  crash
• 15:13
  —
• 15:14
  means
• 15:14
  even
• 15:15
  advanced
• 15:15
  practitioners
• 15:16
  are
• 15:16
  getting
• 15:16
  perhaps
• 15:16
  three
• 15:17
  good
• 15:17
  hours
• 15:17
  of
• 15:17
  deep
• 15:18
  agent-assisted
• 15:18
  work
• 15:18
  per
• 15:19
  day.
• 15:19
  His
• 15:19
  prediction:
• 15:20
  big
• 15:21
  tech
• 15:21
  companies
• 15:21
  are
• 15:21
  quietly
• 15:22
  dying
• 15:22
  because
• 15:23
  their
• 15:23
  organizational
• 15:24
  structures
• 15:24
  cannot
• 15:24
  absorb
• 15:25
  the
• 15:25
  speed,
• 15:26
  and
• 15:26
  small
• 15:26
  teams
• 15:26
  of
• 15:26
  2
• 15:27
  to
• 15:27
  20
• 15:27
  people
• 15:28
  will
• 15:28
  rival
• 15:28
  their
• 15:28
  output.
• 15:29
  The
• 15:29
  open-source
• 15:30
  Gas
• 15:30
  Town
• 15:30
  orchestrator
• 15:31
  he
• 15:31
  built
• 15:31
  reflects
• 15:32
  this
• 15:32
  thesis
• 15:32
  —
• 15:33
  agent
• 15:33
  infrastructure
• 15:34
  designed
• 15:34
  for
• 15:34
  small,
• 15:35
  fast
• 15:35
  teams
• 15:36
  rather
• 15:36
  than
• 15:36
  enterprise
• 15:37
  scale.
• 15:38
  A
• 15:38
  thoughtful
• 15:38
  video
• 15:38
  essay
• 15:39
  documents
• 15:39
  the
• 15:39
  collapse
• 15:40
  of
• 15:40
  the
• 15:40
  knowledge
• 15:40
  ecosystem
• 15:41
  that
• 15:41
  LLMs
• 15:42
  depend
• 15:42
  on.
• 15:43
  Stack
• 15:43
  Overflow
• 15:44
  question
• 15:44
  volume
• 15:45
  dropped
• 15:45
  78%.
• 15:47
  Chegg's
• 15:47
  stock
• 15:47
  fell
• 15:47
  99%.
• 15:49
  Publishers
• 15:49
  lost
• 15:50
  a
• 15:50
  third
• 15:50
  of
• 15:50
  their
• 15:50
  search
• 15:51
  traffic
• 15:51
  to
• 15:51
  AI
• 15:51
  summaries.
• 15:52
  The
• 15:52
  deeper
• 15:53
  problem:
• 15:53
  when
• 15:54
  AI
• 15:54
  models
• 15:54
  train
• 15:54
  on
• 15:55
  content
• 15:55
  generated
• 15:56
  by
• 15:56
  other
• 15:56
  AI
• 15:56
  models,
• 15:57
  output
• 15:57
  quality
• 15:58
  degrades
• 15:58
  —
• 15:59
  the
• 15:59
  model
• 15:59
  collapse
• 15:59
  thesis.
• 16:01
  The
• 16:01
  internet
• 16:01
  is
• 16:01
  filling
• 16:02
  with
• 16:02
  exactly
• 16:02
  the
• 16:02
  kind
• 16:03
  of
• 16:03
  synthetic
• 16:03
  content
• 16:04
  that
• 16:04
  causes
• 16:04
  this
• 16:05
  degradation,
• 16:06
  while
• 16:06
  the
• 16:06
  economic
• 16:07
  incentives
• 16:07
  that
• 16:07
  produced
• 16:08
  high-quality
• 16:08
  human-generated
• 16:09
  training
• 16:09
  data
• 16:10
  (professional
• 16:10
  reputation,
• 16:11
  ad
• 16:11
  revenue,
• 16:12
  community
• 16:13
  engagement)
• 16:13
  are
• 16:13
  evaporating.
• 16:15
  AI
• 16:15
  may
• 16:15
  be
• 16:15
  consuming
• 16:16
  the
• 16:16
  ecosystem
• 16:16
  it
• 16:17
  needs
• 16:17
  to
• 16:17
  survive.
• 16:18
  On
• 16:18
  a
• 16:18
  more
• 16:19
  creative
• 16:19
  note,
• 16:20
  Tencent's
• 16:20
  LeVo
• 16:20
  2
• 16:21
  (SongGeneration
• 16:21
  2)
• 16:22
  is
• 16:22
  an
• 16:22
  open-source
• 16:23
  music
• 16:23
  foundation
• 16:23
  model
• 16:24
  aiming
• 16:24
  for
• 16:24
  commercial-grade
• 16:25
  generation
• 16:25
  quality.
• 16:27
  Early
• 16:27
  community
• 16:27
  reception
• 16:28
  is
• 16:28
  mixed
• 16:28
  —
• 16:29
  orchestral
• 16:29
  generation
• 16:30
  shows
• 16:30
  promise,
• 16:31
  but
• 16:31
  control
• 16:32
  over
• 16:32
  genre
• 16:32
  and
• 16:32
  instrumentation
• 16:33
  remains
• 16:34
  limited
• 16:34
  compared
• 16:34
  to
• 16:34
  the
• 16:34
  description,
• 16:35
  and
• 16:36
  the
• 16:36
  HuggingFace
• 16:36
  demo
• 16:36
  queue
• 16:37
  is
• 16:37
  perpetually
• 16:38
  long.
• 16:38
  For
• 16:38
  anyone
• 16:39
  who
• 16:39
  spent
• 16:39
  time
• 16:39
  with
• 16:39
  Udio
• 16:40
  before
• 16:40
  its
• 16:40
  pricing
• 16:40
  changes,
• 16:41
  the
• 16:41
  appeal
• 16:42
  of
• 16:42
  a
• 16:42
  locally-runnable
• 16:43
  music
• 16:43
  model
• 16:43
  where
• 16:43
  you
• 16:44
  own
• 16:44
  the
• 16:44
  outputs
• 16:44
  is
• 16:44
  clear,
• 16:45
  even
• 16:46
  if
• 16:46
  the
• 16:46
  quality
• 16:46
  gap
• 16:46
  has
• 16:47
  not
• 16:47
  fully
• 16:47
  closed.