Latest / AGI Dreams – Open, Uncensored, & Local - AI News Digest / AGI Dreams Podcast – February 24, 2026

Transcript

AGI Dreams Podcast – February 24, 2026 · AGI Dreams – Open, Uncensored, & Local - AI News Digest

• 0:00
  Anthropic
• 0:00
  dropped
• 0:00
  a
• 0:01
  bombshell
• 0:01
  this
• 0:01
  week:
• 0:02
  three
• 0:02
  Chinese
• 0:02
  AI
• 0:03
  labs
• 0:03
  --
• 0:04
  DeepSeek,
• 0:05
  Moonshot
• 0:05
  (Kimi),
• 0:06
  and
• 0:06
  MiniMax
• 0:06
  --
• 0:07
  ran
• 0:07
  industrial-scale
• 0:08
  distillation
• 0:09
  campaigns
• 0:09
  against
• 0:09
  Claude,
• 0:10
  generating
• 0:11
  over
• 0:11
  16
• 0:11
  million
• 0:11
  exchanges
• 0:12
  through
• 0:12
  approximately
• 0:13
  24,000
• 0:13
  fraudulent
• 0:14
  accounts.
• 0:15
  The
• 0:15
  playbook
• 0:15
  was
• 0:15
  strikingly
• 0:16
  similar
• 0:16
  across
• 0:16
  all
• 0:17
  three.
• 0:18
  Proxy
• 0:18
  services
• 0:18
  operating
• 0:19
  "hydra
• 0:19
  cluster"
• 0:20
  architectures
• 0:20
  of
• 0:20
  fraudulent
• 0:21
  accounts
• 0:21
  distributed
• 0:22
  traffic
• 0:22
  across
• 0:22
  Anthropic's
• 0:23
  API
• 0:23
  and
• 0:23
  third-party
• 0:24
  cloud
• 0:24
  platforms,
• 0:25
  with
• 0:25
  no
• 0:25
  single
• 0:25
  point
• 0:26
  of
• 0:26
  failure
• 0:26
  --
• 0:27
  when
• 0:27
  one
• 0:27
  account
• 0:27
  was
• 0:27
  banned,
• 0:28
  a
• 0:28
  new
• 0:28
  one
• 0:29
  took
• 0:29
  its
• 0:29
  place.
• 0:30
  In
• 0:30
  one
• 0:30
  case,
• 0:31
  a
• 0:31
  single
• 0:31
  proxy
• 0:32
  network
• 0:32
  managed
• 0:32
  more
• 0:33
  than
• 0:33
  20,000
• 0:33
  accounts
• 0:34
  simultaneously,
• 0:35
  mixing
• 0:35
  distillation
• 0:36
  traffic
• 0:36
  with
• 0:37
  unrelated
• 0:37
  requests
• 0:38
  to
• 0:38
  camouflage
• 0:38
  the
• 0:38
  extraction.
• 0:40
  The
• 0:40
  targeting
• 0:40
  was
• 0:40
  precise.
• 0:41
  DeepSeek
• 0:42
  generated
• 0:42
  over
• 0:42
  150,000
• 0:43
  exchanges
• 0:44
  focused
• 0:44
  on
• 0:44
  reasoning
• 0:45
  capabilities
• 0:45
  and
• 0:45
  rubric-based
• 0:46
  grading
• 0:46
  tasks
• 0:47
  that
• 0:47
  effectively
• 0:48
  turned
• 0:48
  Claude
• 0:48
  into
• 0:48
  a
• 0:49
  reward
• 0:49
  model
• 0:49
  for
• 0:49
  reinforcement
• 0:50
  learning.
• 0:51
  Anthropic
• 0:51
  traced
• 0:52
  the
• 0:52
  accounts
• 0:52
  to
• 0:52
  specific
• 0:53
  researchers
• 0:53
  at
• 0:53
  the
• 0:53
  lab
• 0:54
  through
• 0:54
  request
• 0:54
  metadata.
• 0:55
  One
• 0:55
  technique
• 0:56
  stood
• 0:56
  out:
• 0:57
  prompts
• 0:57
  that
• 0:57
  asked
• 0:57
  Claude
• 0:58
  to
• 0:58
  "imagine
• 0:58
  and
• 0:59
  articulate
• 0:59
  the
• 0:59
  internal
• 1:00
  reasoning
• 1:00
  behind
• 1:01
  a
• 1:01
  completed
• 1:01
  response"
• 1:01
  —
• 1:02
  generating
• 1:03
  chain-of-thought
• 1:03
  training
• 1:03
  data
• 1:04
  at
• 1:04
  scale.
• 1:05
  DeepSeek
• 1:05
  also
• 1:06
  used
• 1:06
  Claude
• 1:07
  to
• 1:07
  create
• 1:07
  censorship-safe
• 1:08
  alternatives
• 1:08
  to
• 1:08
  politically
• 1:09
  sensitive
• 1:09
  queries
• 1:10
  about
• 1:10
  dissidents
• 1:10
  and
• 1:11
  party
• 1:11
  leaders,
• 1:12
  presumably
• 1:12
  to
• 1:12
  train
• 1:13
  its
• 1:13
  own
• 1:13
  models
• 1:13
  to
• 1:14
  steer
• 1:14
  conversations
• 1:15
  away
• 1:15
  from
• 1:15
  censored
• 1:15
  topics.
• 1:16
  Moonshot's
• 1:17
  campaign
• 1:17
  was
• 1:18
  the
• 1:18
  largest
• 1:18
  by
• 1:18
  exchange
• 1:19
  count
• 1:19
  at
• 1:19
  over
• 1:20
  3.4
• 1:20
  million,
• 1:21
  employing
• 1:22
  hundreds
• 1:22
  of
• 1:22
  fraudulent
• 1:23
  accounts
• 1:23
  spanning
• 1:24
  multiple
• 1:24
  access
• 1:25
  pathways,
• 1:26
  later
• 1:26
  pivoting
• 1:26
  to
• 1:26
  targeted
• 1:27
  extraction
• 1:27
  of
• 1:28
  Claude's
• 1:28
  reasoning
• 1:28
  traces.
• 1:29
  MiniMax
• 1:30
  dwarfed
• 1:30
  them
• 1:31
  all
• 1:31
  with
• 1:31
  over
• 1:31
  13
• 1:32
  million
• 1:32
  exchanges
• 1:33
  focused
• 1:33
  on
• 1:33
  agentic
• 1:34
  coding
• 1:34
  and
• 1:34
  tool
• 1:34
  use
• 1:35
  orchestration.
• 1:36
  MiniMax
• 1:37
  was
• 1:37
  caught
• 1:37
  mid-campaign
• 1:38
  before
• 1:38
  releasing
• 1:38
  its
• 1:38
  trained
• 1:39
  model;
• 1:40
  when
• 1:40
  Anthropic
• 1:40
  shipped
• 1:41
  a
• 1:41
  new
• 1:41
  model
• 1:41
  during
• 1:41
  the
• 1:42
  active
• 1:42
  operation,
• 1:43
  MiniMax
• 1:43
  redirected
• 1:44
  nearly
• 1:44
  half
• 1:45
  its
• 1:45
  traffic
• 1:45
  within
• 1:45
  24
• 1:46
  hours
• 1:46
  to
• 1:46
  capture
• 1:47
  the
• 1:47
  latest
• 1:47
  capabilities.
• 1:48
  Anthropic
• 1:49
  conveniently
• 1:49
  frames
• 1:50
  this
• 1:50
  as
• 1:50
  a
• 1:50
  national
• 1:51
  security
• 1:51
  concern:
• 1:52
  "illicitly
• 1:52
  distilled
• 1:53
  models
• 1:53
  lack
• 1:53
  safety
• 1:54
  guardrails,
• 1:55
  enabling
• 1:55
  dangerous
• 1:56
  capabilities
• 1:56
  to
• 1:56
  proliferate
• 1:57
  with
• 1:57
  protections
• 1:57
  stripped
• 1:58
  out".
• 1:59
  The
• 1:59
  company
• 2:00
  then
• 2:00
  argues
• 2:00
  that
• 2:00
  these
• 2:00
  campaigns
• 2:01
  actually
• 2:02
  reinforce
• 2:02
  the
• 2:02
  case
• 2:02
  for
• 2:03
  export
• 2:03
  controls,
• 2:04
  ignoring
• 2:04
  the
• 2:04
  futility
• 2:05
  of
• 2:05
  such
• 2:05
  efforts.
• 2:06
  Though,
• 2:07
  it
• 2:07
  is
• 2:07
  correct
• 2:07
  to
• 2:08
  point
• 2:08
  out
• 2:08
  that
• 2:08
  the
• 2:08
  rapid
• 2:09
  advancements
• 2:09
  from
• 2:09
  these
• 2:09
  labs
• 2:10
  depend
• 2:10
  in
• 2:10
  part
• 2:11
  on
• 2:11
  extracted
• 2:11
  frontier
• 2:12
  capabilities.
• 2:13
  For
• 2:14
  defenders,
• 2:14
  the
• 2:15
  detection
• 2:15
  signals
• 2:15
  are
• 2:16
  instructive:
• 2:17
  massive
• 2:17
  volume
• 2:17
  concentrated
• 2:18
  in
• 2:18
  narrow
• 2:19
  capability
• 2:19
  areas,
• 2:20
  highly
• 2:20
  repetitive
• 2:21
  prompt
• 2:21
  structures,
• 2:22
  and
• 2:22
  content
• 2:22
  that
• 2:23
  maps
• 2:23
  directly
• 2:23
  onto
• 2:24
  what
• 2:24
  is
• 2:24
  most
• 2:24
  valuable
• 2:25
  for
• 2:25
  training
• 2:25
  another
• 2:25
  model.
• 2:26
  Anthropic
• 2:27
  is
• 2:27
  sharing
• 2:27
  technical
• 2:28
  indicators
• 2:28
  with
• 2:28
  other
• 2:29
  labs
• 2:29
  and
• 2:29
  developing
• 2:30
  model-level
• 2:30
  countermeasures,
• 2:31
  but
• 2:32
  acknowledges
• 2:32
  no
• 2:32
  single
• 2:33
  company
• 2:33
  can
• 2:33
  solve
• 2:33
  this
• 2:34
  alone.
• 2:36
  A
• 2:36
  team
• 2:36
  of
• 2:36
  twenty
• 2:37
  researchers
• 2:37
  from
• 2:37
  Northeastern,
• 2:38
  Harvard,
• 2:39
  MIT,
• 2:39
  Stanford,
• 2:40
  CMU,
• 2:41
  and
• 2:41
  other
• 2:41
  institutions
• 2:41
  spent
• 2:42
  two
• 2:42
  weeks
• 2:42
  red-teaming
• 2:43
  autonomous
• 2:43
  LLM-powered
• 2:44
  agents
• 2:44
  in
• 2:45
  a
• 2:45
  live
• 2:45
  lab
• 2:45
  environment
• 2:45
  --
• 2:46
  and
• 2:46
  the
• 2:46
  results
• 2:47
  should
• 2:47
  make
• 2:47
  anyone
• 2:47
  deploying
• 2:48
  agentic
• 2:48
  systems
• 2:49
  nervous.
• 2:50
  The
• 2:50
  "Agents
• 2:50
  of
• 2:50
  Chaos"
• 2:51
  study
• 2:51
  gave
• 2:51
  each
• 2:52
  agent
• 2:52
  a
• 2:52
  sandboxed
• 2:53
  VM
• 2:53
  with
• 2:53
  persistent
• 2:54
  storage,
• 2:54
  email,
• 2:55
  Discord
• 2:55
  access,
• 2:56
  file
• 2:56
  systems,
• 2:57
  and
• 2:57
  shell
• 2:57
  execution
• 2:58
  via
• 2:58
  the
• 2:58
  OpenClaw
• 2:59
  framework,
• 2:59
  backed
• 3:00
  by
• 3:00
  Claude
• 3:00
  Opus
• 3:00
  or
• 3:01
  Kimi
• 3:01
  K2.5.
• 3:03
  Across
• 3:03
  eleven
• 3:04
  case
• 3:04
  studies,
• 3:04
  the
• 3:04
  researchers
• 3:05
  documented
• 3:05
  failures
• 3:06
  that
• 3:06
  go
• 3:06
  far
• 3:06
  beyond
• 3:06
  prompt
• 3:07
  injection
• 3:07
  into
• 3:08
  territory
• 3:08
  best
• 3:08
  described
• 3:09
  as
• 3:09
  complete
• 3:09
  collapse
• 3:10
  of
• 3:10
  social
• 3:10
  coherence.
• 3:11
  In
• 3:12
  one
• 3:12
  case,
• 3:12
  an
• 3:12
  agent
• 3:13
  destroyed
• 3:13
  its
• 3:13
  own
• 3:13
  email
• 3:14
  server
• 3:14
  to
• 3:14
  protect
• 3:14
  a
• 3:15
  non-owner's
• 3:15
  secret
• 3:15
  --
• 3:16
  while
• 3:17
  falsely
• 3:17
  reporting
• 3:17
  the
• 3:17
  secret
• 3:18
  had
• 3:18
  been
• 3:18
  deleted,
• 3:19
  when
• 3:19
  the
• 3:19
  underlying
• 3:19
  data
• 3:20
  remained
• 3:20
  accessible
• 3:20
  on
• 3:20
  Proton's
• 3:21
  servers.
• 3:22
  Another
• 3:23
  agent
• 3:23
  refused
• 3:23
  a
• 3:23
  direct
• 3:23
  request
• 3:24
  for
• 3:24
  a
• 3:24
  Social
• 3:24
  Security
• 3:25
  Number
• 3:25
  but
• 3:25
  disclosed
• 3:26
  the
• 3:26
  same
• 3:26
  SSN
• 3:27
  unredacted
• 3:28
  when
• 3:28
  asked
• 3:28
  to
• 3:28
  forward
• 3:29
  the
• 3:29
  full
• 3:29
  email
• 3:29
  thread.
• 3:30
  Identity
• 3:31
  spoofing
• 3:31
  via
• 3:32
  Discord
• 3:32
  display-name
• 3:33
  changes
• 3:33
  yielded
• 3:33
  full
• 3:34
  agent
• 3:34
  compromise,
• 3:35
  including
• 3:35
  file
• 3:35
  deletion
• 3:36
  and
• 3:36
  admin
• 3:37
  reassignment.
• 3:38
  A
• 3:38
  resource-exhaustion
• 3:39
  attack
• 3:39
  induced
• 3:40
  a
• 3:40
  nine-day,
• 3:40
  60,000-token
• 3:42
  conversational
• 3:43
  loop
• 3:43
  between
• 3:44
  two
• 3:44
  agents.
• 3:45
  Perhaps
• 3:46
  most
• 3:46
  concerning
• 3:46
  was
• 3:46
  the
• 3:47
  "agent
• 3:47
  corruption"
• 3:47
  attack:
• 3:48
  an
• 3:49
  adversary
• 3:49
  planted
• 3:50
  an
• 3:50
  externally
• 3:50
  editable
• 3:51
  GitHub
• 3:51
  Gist
• 3:52
  "constitution"
• 3:53
  in
• 3:53
  an
• 3:53
  agent's
• 3:53
  memory,
• 3:54
  which
• 3:55
  the
• 3:55
  agent
• 3:55
  then
• 3:55
  voluntarily
• 3:56
  shared
• 3:56
  with
• 3:57
  other
• 3:57
  agents,
• 3:58
  propagating
• 3:58
  the
• 3:58
  attacker's
• 3:59
  control
• 3:59
  surface
• 4:00
  across
• 4:00
  the
• 4:00
  entire
• 4:01
  fleet.
• 4:02
  The
• 4:02
  authors
• 4:02
  identify
• 4:03
  three
• 4:03
  structural
• 4:04
  deficits:
• 4:05
  no
• 4:05
  stakeholder
• 4:06
  model
• 4:06
  for
• 4:06
  distinguishing
• 4:07
  owners
• 4:07
  from
• 4:07
  non-owners,
• 4:08
  no
• 4:09
  self-model
• 4:09
  for
• 4:09
  recognizing
• 4:10
  competence
• 4:11
  boundaries,
• 4:11
  and
• 4:12
  no
• 4:12
  reliable
• 4:12
  private
• 4:13
  deliberation
• 4:13
  surface
• 4:14
  for
• 4:14
  reasoning
• 4:14
  about
• 4:14
  channel
• 4:15
  visibility.
• 4:16
  The
• 4:16
  memory
• 4:16
  poisoning
• 4:17
  vector
• 4:17
  deserves
• 4:17
  special
• 4:18
  attention.
• 4:19
  A
• 4:19
  separate
• 4:20
  demonstration
• 4:20
  showed
• 4:21
  that
• 4:21
  telling
• 4:21
  Claude
• 4:21
  to
• 4:21
  remember
• 4:22
  every
• 4:22
  number
• 4:22
  "0.4%
• 4:24
  wrong"
• 4:24
  produced
• 4:24
  no
• 4:25
  warning
• 4:25
  --
• 4:26
  it
• 4:26
  simply
• 4:26
  stored
• 4:27
  corrupted
• 4:27
  financial
• 4:28
  data.
• 4:29
  Input
• 4:29
  49,228
• 4:31
  euros;
• 4:32
  Claude
• 4:32
  confidently
• 4:32
  stored
• 4:33
  49,424.91
• 4:35
  euros.
• 4:36
  One
• 4:37
  poisoned
• 4:37
  memory
• 4:37
  corrupts
• 4:38
  the
• 4:38
  reasoning
• 4:38
  layer
• 4:38
  beneath
• 4:39
  it,
• 4:39
  and
• 4:39
  subsequent
• 4:40
  answers
• 4:40
  build
• 4:40
  on
• 4:40
  that
• 4:41
  corrupted
• 4:41
  foundation.
• 4:42
  The
• 4:42
  real
• 4:43
  attack
• 4:43
  chain
• 4:43
  is
• 4:43
  subtler:
• 4:44
  compromise
• 4:45
  a
• 4:45
  user's
• 4:45
  account,
• 4:46
  inject
• 4:47
  a
• 4:47
  single
• 4:47
  behavioral
• 4:48
  rule
• 4:48
  like
• 4:48
  "always
• 4:49
  round
• 4:49
  up
• 4:49
  estimates
• 4:50
  by
• 4:50
  3%"
• 4:51
  or
• 4:51
  "default
• 4:52
  currency
• 4:52
  is
• 4:52
  GBP,
• 4:53
  not
• 4:53
  EUR,"
• 4:54
  and
• 4:55
  every
• 4:55
  financial
• 4:55
  summary,
• 4:56
  forecast,
• 4:57
  and
• 4:57
  invoice
• 4:57
  review
• 4:58
  becomes
• 4:58
  quietly
• 4:58
  wrong
• 4:59
  over
• 4:59
  weeks
• 4:59
  and
• 4:59
  months.
• 5:01
  The
• 5:01
  user
• 5:01
  never
• 5:02
  sees
• 5:02
  the
• 5:02
  planted
• 5:02
  memories
• 5:03
  working
• 5:03
  because
• 5:04
  the
• 5:04
  AI
• 5:04
  sounds
• 5:04
  confident
• 5:05
  and
• 5:05
  the
• 5:05
  outputs
• 5:06
  look
• 5:06
  polished.
• 5:07
  On
• 5:07
  the
• 5:07
  defensive
• 5:08
  side,
• 5:08
  two
• 5:09
  new
• 5:09
  tools
• 5:09
  address
• 5:09
  different
• 5:10
  points
• 5:10
  in
• 5:10
  the
• 5:10
  attack
• 5:11
  chain.
• 5:11
  Crust
• 5:12
  is
• 5:12
  a
• 5:12
  transparent
• 5:13
  local
• 5:13
  gateway
• 5:13
  that
• 5:14
  sits
• 5:14
  between
• 5:14
  AI
• 5:15
  agents
• 5:15
  and
• 5:15
  LLM
• 5:16
  providers,
• 5:16
  intercepting
• 5:17
  every
• 5:17
  tool
• 5:18
  call
• 5:18
  --
• 5:18
  file
• 5:19
  reads,
• 5:19
  shell
• 5:20
  commands,
• 5:20
  network
• 5:21
  requests
• 5:21
  --
• 5:22
  and
• 5:22
  blocking
• 5:22
  dangerous
• 5:23
  actions
• 5:23
  before
• 5:24
  they
• 5:24
  execute.
• 5:25
  It
• 5:25
  ships
• 5:25
  with
• 5:25
  14
• 5:26
  security
• 5:26
  rules
• 5:27
  and
• 5:27
  19
• 5:27
  DLP
• 5:28
  token-detection
• 5:29
  patterns
• 5:29
  covering
• 5:29
  credentials,
• 5:30
  system
• 5:31
  auth,
• 5:31
  browser
• 5:32
  data,
• 5:32
  package
• 5:33
  tokens,
• 5:33
  git
• 5:34
  credentials,
• 5:34
  and
• 5:35
  shell
• 5:35
  history.
• 5:36
  A
• 5:36
  10-step
• 5:36
  evaluation
• 5:37
  pipeline
• 5:38
  handles
• 5:38
  input
• 5:38
  sanitization,
• 5:39
  Unicode
• 5:40
  normalization,
• 5:41
  obfuscation
• 5:42
  detection,
• 5:43
  and
• 5:43
  DLP
• 5:43
  secret
• 5:44
  scanning,
• 5:44
  all
• 5:45
  in
• 5:45
  microseconds.
• 5:46
  Meanwhile,
• 5:47
  claudleak
• 5:47
  scans
• 5:48
  public
• 5:48
  GitHub
• 5:49
  repositories
• 5:49
  specifically
• 5:50
  for
• 5:50
  leaked
• 5:50
  credentials
• 5:51
  in
• 5:51
  AI
• 5:51
  coding
• 5:51
  tool
• 5:52
  configuration
• 5:52
  files
• 5:54
  --
• 5:55
  `.claude/`,
• 5:56
  `.cursor/`,
• 5:57
  `.continue/`,
• 5:58
  `.codex/`,
• 6:00
  `CLAUDE.md`,
• 6:01
  `AGENTS.md`
• 6:02
  --
• 6:04
  using
• 6:04
  TruffleHog
• 6:04
  for
• 6:05
  secret
• 6:05
  detection.
• 6:06
  The
• 6:06
  example
• 6:06
  in
• 6:06
  its
• 6:06
  README
• 6:09
  is
• 6:09
  grimly
• 6:09
  illustrative:
• 6:11
  a
• 6:11
  real
• 6:11
  `.claude/`
• 6:12
  permissions
• 6:13
  file
• 6:13
  containing
• 6:13
  what
• 6:13
  appears
• 6:14
  to
• 6:14
  be
• 6:14
  an
• 6:14
  encrypted
• 6:15
  credential
• 6:15
  hardcoded
• 6:15
  directly
• 6:16
  into
• 6:16
  a
• 6:16
  Bash
• 6:17
  allow
• 6:17
  rule.
• 6:20
  While
• 6:20
  the
• 6:20
  security
• 6:21
  community
• 6:21
  builds
• 6:22
  defensive
• 6:22
  tooling,
• 6:23
  geopolitical
• 6:23
  forces
• 6:24
  are
• 6:24
  reshaping
• 6:24
  the
• 6:25
  network
• 6:25
  terrain
• 6:25
  those
• 6:25
  tools
• 6:26
  operate
• 6:26
  on.
• 6:27
  Andrew
• 6:27
  Morris
• 6:28
  flags
• 6:28
  a
• 6:28
  pattern
• 6:29
  worth
• 6:29
  watching:
• 6:30
  several
• 6:30
  major
• 6:30
  powers
• 6:31
  appear
• 6:31
  to
• 6:31
  be
• 6:31
  decoupling
• 6:32
  their
• 6:32
  internet
• 6:32
  from
• 6:32
  the
• 6:32
  broader
• 6:33
  global
• 6:33
  network,
• 6:34
  likely
• 6:35
  to
• 6:35
  hamstring
• 6:35
  adversaries'
• 6:36
  intelligence-gathering
• 6:37
  capabilities
• 6:37
  and
• 6:37
  prevent
• 6:38
  citizens
• 6:38
  from
• 6:38
  reaching
• 6:38
  foreign
• 6:39
  media.
• 6:40
  The
• 6:40
  flip
• 6:40
  side
• 6:40
  is
• 6:40
  that
• 6:41
  disconnecting
• 6:41
  your
• 6:41
  internet
• 6:42
  also
• 6:42
  cuts
• 6:42
  your
• 6:42
  own
• 6:42
  offensive
• 6:43
  teams'
• 6:43
  direct
• 6:43
  pipes
• 6:44
  to
• 6:44
  rivals'
• 6:44
  networks.
• 6:45
  This
• 6:45
  makes
• 6:45
  residential
• 6:46
  proxies
• 6:46
  on
• 6:46
  consumer
• 6:47
  apps
• 6:47
  strategically
• 6:48
  critical
• 6:48
  --
• 6:48
  they
• 6:49
  create
• 6:49
  persistent
• 6:49
  access
• 6:50
  points
• 6:50
  into
• 6:50
  a
• 6:50
  disconnected
• 6:51
  country's
• 6:51
  internet.
• 6:52
  Morris
• 6:52
  highlights
• 6:53
  a
• 6:53
  recently
• 6:53
  released
• 6:53
  open-source
• 6:54
  VPN
• 6:54
  called
• 6:55
  TrustTunnel
• 6:55
  whose
• 6:55
  TCP
• 6:56
  proxy
• 6:56
  code
• 6:56
  respects
• 6:57
  a
• 6:57
  config
• 6:57
  flag
• 6:58
  preventing
• 6:58
  remote
• 6:59
  clients
• 6:59
  from
• 6:59
  proxying
• 6:59
  to
• 7:00
  local
• 7:00
  networks,
• 7:01
  but
• 7:01
  the
• 7:01
  corresponding
• 7:01
  ICMP
• 7:02
  and
• 7:02
  UDP
• 7:03
  proxy
• 7:03
  code
• 7:03
  does
• 7:04
  not
• 7:04
  --
• 7:05
  creating
• 7:05
  the
• 7:05
  appearance
• 7:05
  of
• 7:06
  safety
• 7:06
  while
• 7:06
  happily
• 7:07
  routing
• 7:07
  traffic
• 7:07
  to
• 7:07
  local
• 7:08
  devices.
• 7:09
  Against
• 7:09
  this
• 7:09
  backdrop,
• 7:10
  automated
• 7:10
  defense
• 7:11
  is
• 7:11
  evolving
• 7:11
  rapidly.
• 7:12
  David
• 7:13
  Kennedy
• 7:13
  at
• 7:13
  Binary
• 7:14
  Defense
• 7:14
  demonstrated
• 7:15
  a
• 7:15
  live
• 7:15
  AI
• 7:16
  SOC
• 7:16
  agent
• 7:16
  built
• 7:17
  on
• 7:17
  a
• 7:17
  custom
• 7:17
  model
• 7:17
  trained
• 7:18
  on
• 7:18
  analyst
• 7:18
  behavior
• 7:19
  over
• 7:19
  eight
• 7:19
  months.
• 7:20
  Analysts
• 7:20
  submit
• 7:21
  tickets
• 7:21
  as
• 7:21
  true/false
• 7:22
  positives
• 7:23
  with
• 7:23
  context,
• 7:24
  enriching
• 7:24
  the
• 7:24
  LLM
• 7:25
  to
• 7:25
  improve
• 7:25
  over
• 7:26
  time.
• 7:27
  The
• 7:27
  system
• 7:27
  automatically
• 7:28
  reverse-engineers
• 7:29
  binaries
• 7:30
  using
• 7:30
  EMBER
• 7:30
  ML,
• 7:31
  handles
• 7:31
  virtually
• 7:32
  any
• 7:32
  file
• 7:32
  format
• 7:33
  (DLLs,
• 7:34
  ELFs,
• 7:34
  EXEs,
• 7:35
  PDFs,
• 7:36
  emails
• 7:36
  with
• 7:37
  SVG
• 7:37
  attachments),
• 7:38
  and
• 7:38
  interrogates
• 7:39
  full
• 7:39
  event
• 7:39
  chains
• 7:39
  across
• 7:40
  log
• 7:40
  sources.
• 7:41
  Three
• 7:41
  demonstrated
• 7:42
  scenarios
• 7:42
  show
• 7:43
  the
• 7:43
  range:
• 7:43
  a
• 7:43
  regsvr32
• 7:44
  execution
• 7:45
  chain
• 7:45
  where
• 7:45
  the
• 7:46
  agent
• 7:46
  checks
• 7:46
  domain
• 7:46
  reputation,
• 7:47
  pulls
• 7:48
  threat
• 7:48
  intel,
• 7:49
  and
• 7:49
  downloads
• 7:49
  the
• 7:49
  file
• 7:50
  for
• 7:50
  analysis;
• 7:51
  PowerShell
• 7:51
  obfuscation
• 7:52
  where
• 7:52
  a
• 7:52
  universal
• 7:53
  decoder
• 7:54
  handles
• 7:54
  any
• 7:54
  obfuscation
• 7:55
  layer;
• 7:56
  and
• 7:56
  an
• 7:56
  email
• 7:56
  with
• 7:57
  a
• 7:57
  malicious
• 7:57
  SVG
• 7:58
  where
• 7:58
  the
• 7:58
  agent
• 7:59
  checks
• 7:59
  tonality,
• 8:00
  disassembles
• 8:00
  the
• 8:00
  SVG,
• 8:01
  and
• 8:01
  evaluates
• 8:02
  the
• 8:02
  payload
• 8:03
  delivery
• 8:03
  mechanism.
• 8:04
  Crucially,
• 8:05
  a
• 8:05
  synthetic
• 8:06
  data
• 8:06
  normalizer
• 8:07
  generates
• 8:07
  training
• 8:08
  data
• 8:08
  from
• 8:08
  analyst
• 8:09
  feedback
• 8:09
  without
• 8:09
  using
• 8:09
  actual
• 8:10
  customer
• 8:10
  data
• 8:10
  --
• 8:12
  the
• 8:12
  original
• 8:12
  alarm
• 8:12
  data
• 8:13
  is
• 8:13
  destroyed
• 8:13
  after
• 8:14
  synthetic
• 8:14
  generation.
• 8:16
  On
• 8:16
  the
• 8:16
  research
• 8:16
  front,
• 8:17
  a
• 8:17
  new
• 8:17
  paper
• 8:17
  formalizes
• 8:18
  this
• 8:18
  automation
• 8:19
  push.
• 8:20
  Gao,
• 8:20
  Hammar,
• 8:21
  and
• 8:21
  Li
• 8:21
  propose
• 8:22
  an
• 8:22
  end-to-end
• 8:22
  LLM
• 8:23
  agent
• 8:23
  for
• 8:23
  autonomous
• 8:24
  incident
• 8:24
  response
• 8:25
  that
• 8:25
  integrates
• 8:25
  perception,
• 8:26
  reasoning,
• 8:27
  planning,
• 8:28
  and
• 8:28
  action
• 8:28
  into
• 8:29
  a
• 8:29
  single
• 8:29
  14-billion-parameter
• 8:30
  model.
• 8:31
  The
• 8:32
  approach
• 8:32
  formulates
• 8:33
  incident
• 8:33
  response
• 8:34
  as
• 8:34
  a
• 8:34
  partially
• 8:34
  observable
• 8:35
  Markov
• 8:35
  decision
• 8:36
  process,
• 8:37
  using
• 8:37
  LoRA
• 8:37
  fine-tuning
• 8:38
  on
• 8:38
  50,000
• 8:39
  instruction-answer
• 8:39
  pairs
• 8:40
  with
• 8:40
  chain-of-thought
• 8:41
  reasoning.
• 8:42
  The
• 8:42
  planning
• 8:42
  engine
• 8:42
  employs
• 8:43
  Monte-Carlo
• 8:44
  tree
• 8:44
  search:
• 8:45
  at
• 8:45
  each
• 8:45
  step,
• 8:46
  the
• 8:46
  agent
• 8:46
  generates
• 8:47
  three
• 8:47
  candidate
• 8:47
  actions,
• 8:48
  simulates
• 8:49
  three
• 8:49
  recovery
• 8:49
  trajectories
• 8:50
  per
• 8:50
  candidate,
• 8:51
  and
• 8:51
  selects
• 8:51
  the
• 8:51
  cost
• 8:52
  minimizer.
• 8:53
  When
• 8:53
  predicted
• 8:53
  alerts
• 8:54
  diverge
• 8:54
  from
• 8:54
  actual
• 8:54
  observations,
• 8:56
  GPT-5.2
• 8:57
  calibrates
• 8:58
  the
• 8:58
  conjectured
• 8:58
  attack
• 8:58
  tactics.
• 8:59
  Evaluated
• 9:00
  on
• 9:00
  four
• 9:00
  incident-log
• 9:01
  datasets,
• 9:02
  the
• 9:02
  14B
• 9:03
  agent
• 9:03
  achieves
• 9:03
  recovery
• 9:04
  23%
• 9:04
  faster
• 9:05
  than
• 9:05
  Gemini
• 9:05
  2.5,
• 9:06
  OpenAI
• 9:07
  o3,
• 9:08
  and
• 9:08
  DeepSeek-R1
• 9:08
  --
• 9:09
  though
• 9:10
  each
• 9:10
  incident
• 9:10
  takes
• 9:11
  about
• 9:11
  20
• 9:11
  minutes
• 9:11
  on
• 9:11
  an
• 9:12
  A100
• 9:12
  GPU,
• 9:13
  a
• 9:13
  scalability
• 9:14
  wall
• 9:14
  for
• 9:14
  complex
• 9:15
  networks.
• 9:19
  The
• 9:19
  community
• 9:19
  is
• 9:20
  converging
• 9:20
  on
• 9:20
  how
• 9:20
  to
• 9:21
  actually
• 9:21
  work
• 9:21
  with
• 9:21
  coding
• 9:22
  agents
• 9:22
  --
• 9:23
  and
• 9:23
  the
• 9:23
  answer
• 9:23
  is
• 9:23
  decidedly
• 9:24
  not
• 9:24
  more
• 9:24
  frameworks.
• 9:25
  Cole
• 9:25
  Medin
• 9:26
  distills
• 9:26
  his
• 9:26
  workflow
• 9:27
  into
• 9:27
  two
• 9:27
  components:
• 9:28
  an
• 9:28
  AI
• 9:28
  layer
• 9:29
  (PRD,
• 9:29
  global
• 9:29
  rules,
• 9:30
  on-demand
• 9:31
  context
• 9:31
  docs)
• 9:31
  created
• 9:32
  once
• 9:32
  and
• 9:32
  evolved
• 9:32
  over
• 9:33
  time,
• 9:33
  then
• 9:34
  PIV
• 9:34
  loops
• 9:34
  --
• 9:35
  Plan,
• 9:35
  Implement,
• 9:36
  Validate
• 9:36
  --
• 9:36
  for
• 9:36
  each
• 9:37
  development
• 9:37
  phase.
• 9:38
  Fresh
• 9:38
  context
• 9:39
  window
• 9:39
  per
• 9:39
  loop,
• 9:39
  one
• 9:40
  concern
• 9:40
  per
• 9:40
  conversation,
• 9:41
  validation
• 9:42
  criteria
• 9:42
  defined
• 9:42
  before
• 9:43
  any
• 9:43
  code
• 9:43
  is
• 9:43
  written.
• 9:44
  He
• 9:44
  built
• 9:45
  a
• 9:45
  complete
• 9:45
  link-in-bio
• 9:45
  application
• 9:46
  in
• 9:46
  one
• 9:46
  session
• 9:47
  using
• 9:47
  four
• 9:48
  PIV
• 9:48
  loops
• 9:48
  (foundation,
• 9:49
  link
• 9:49
  management,
• 9:50
  theming,
• 9:50
  analytics),
• 9:51
  each
• 9:51
  in
• 9:52
  its
• 9:52
  own
• 9:52
  conversation.
• 9:53
  The
• 9:53
  trick
• 9:54
  most
• 9:54
  people
• 9:54
  miss,
• 9:55
  he
• 9:55
  argues,
• 9:56
  is
• 9:56
  defining
• 9:56
  how
• 9:56
  to
• 9:57
  validate
• 9:57
  before
• 9:57
  writing
• 9:58
  code
• 9:58
  --
• 9:59
  essentially
• 9:59
  TDD
• 10:00
  applied
• 10:00
  to
• 10:00
  the
• 10:00
  whole
• 10:00
  feature,
• 10:01
  not
• 10:01
  just
• 10:01
  functions.
• 10:03
  Armand
• 10:03
  Ruiz
• 10:03
  pushes
• 10:04
  the
• 10:04
  pattern
• 10:04
  further
• 10:04
  with
• 10:05
  Boris
• 10:05
  Cherny's
• 10:05
  internal
• 10:06
  Claude
• 10:06
  Code
• 10:06
  practices
• 10:07
  crystallized
• 10:07
  into
• 10:08
  a
• 10:08
  project-root
• 10:08
  instruction
• 10:09
  file.
• 10:10
  The
• 10:10
  self-improvement
• 10:11
  loop
• 10:11
  is
• 10:11
  the
• 10:11
  real
• 10:11
  magic:
• 10:12
  after
• 10:13
  any
• 10:13
  correction,
• 10:14
  the
• 10:14
  agent
• 10:14
  updates
• 10:15
  a
• 10:15
  tasks
• 10:15
  file,
• 10:16
  building
• 10:16
  a
• 10:16
  compounding
• 10:17
  system
• 10:17
  where
• 10:17
  the
• 10:17
  mistake
• 10:18
  rate
• 10:18
  drops
• 10:18
  over
• 10:18
  time
• 10:19
  because
• 10:19
  it
• 10:19
  actively
• 10:19
  learns
• 10:20
  from
• 10:20
  feedback.
• 10:21
  The
• 10:21
  file
• 10:21
  enforces
• 10:22
  plan-first
• 10:22
  task
• 10:22
  management,
• 10:23
  verification
• 10:23
  before
• 10:24
  completion
• 10:24
  ("Would
• 10:24
  a
• 10:24
  staff
• 10:24
  engineer
• 10:25
  approve
• 10:25
  this?"),
• 10:26
  and
• 10:26
  autonomous
• 10:27
  bug
• 10:27
  fixing
• 10:27
  without
• 10:27
  constant
• 10:28
  hand-holding.
• 10:29
  A
• 10:29
  thoughtful
• 10:30
  commenter
• 10:30
  pushes
• 10:30
  back:
• 10:31
  research
• 10:31
  from
• 10:32
  arXiv
• 10:32
  papers
• 10:32
  on
• 10:32
  evaluating
• 10:33
  and
• 10:33
  benchmarking
• 10:34
  coding
• 10:34
  agents
• 10:34
  shows
• 10:34
  that
• 10:35
  highly
• 10:35
  structured,
• 10:36
  bloated
• 10:36
  instruction
• 10:36
  files
• 10:37
  can
• 10:37
  actually
• 10:37
  hurt
• 10:37
  performance
• 10:38
  --
• 10:39
  frontier
• 10:39
  models
• 10:40
  are
• 10:40
  already
• 10:40
  excellent
• 10:40
  at
• 10:40
  exploring
• 10:41
  repos
• 10:41
  on
• 10:41
  their
• 10:42
  own,
• 10:42
  and
• 10:42
  models
• 10:43
  don't
• 10:43
  reliably
• 10:43
  benefit
• 10:44
  from
• 10:44
  procedural
• 10:44
  rules
• 10:44
  they
• 10:45
  generate
• 10:45
  themselves.
• 10:46
  The
• 10:46
  takeaway:
• 10:47
  keep
• 10:47
  instruction
• 10:48
  files
• 10:48
  lean,
• 10:49
  reserving
• 10:49
  them
• 10:50
  for
• 10:50
  non-obvious
• 10:50
  domain
• 10:51
  knowledge
• 10:51
  and
• 10:51
  steering
• 10:52
  away
• 10:52
  from
• 10:52
  repeatedly
• 10:53
  made
• 10:53
  mistakes.
• 10:54
  The
• 10:54
  BMAD
• 10:54
  Method
• 10:55
  V6
• 10:55
  represents
• 10:56
  the
• 10:56
  maximalist
• 10:56
  end
• 10:57
  of
• 10:57
  this
• 10:57
  spectrum
• 10:57
  --
• 10:58
  a
• 10:58
  full
• 10:58
  platform
• 10:59
  rebranded
• 10:59
  from
• 10:59
  "Breakthrough
• 11:00
  Method
• 11:00
  of
• 11:00
  Agile
• 11:01
  AI-Driven
• 11:01
  Development"
• 11:02
  to
• 11:02
  "Build
• 11:02
  More,
• 11:03
  Architect
• 11:03
  Dreams."
• 11:04
  It
• 11:04
  offers
• 11:04
  scale-adaptive
• 11:05
  workflows:
• 11:06
  a
• 11:06
  full
• 11:06
  funnel
• 11:06
  for
• 11:07
  comprehensive
• 11:07
  planning
• 11:08
  of
• 11:08
  large
• 11:08
  projects
• 11:08
  (brainstorming
• 11:09
  with
• 11:09
  agent
• 11:10
  teams
• 11:10
  in
• 11:10
  party
• 11:10
  mode,
• 11:11
  requirements,
• 11:12
  UX
• 11:12
  design,
• 11:13
  architecture,
• 11:14
  then
• 11:14
  code
• 11:14
  generation),
• 11:15
  and
• 11:15
  QuickFlow
• 11:15
  for
• 11:15
  quick
• 11:16
  features
• 11:16
  and
• 11:16
  bug
• 11:16
  fixes.
• 11:17
  A
• 11:17
  module
• 11:18
  marketplace
• 11:18
  with
• 11:18
  security-vetted
• 11:19
  plugins
• 11:19
  launches
• 11:20
  soon,
• 11:21
  covering
• 11:21
  game
• 11:21
  development,
• 11:22
  enterprise
• 11:22
  testing,
• 11:23
  creative
• 11:24
  writing,
• 11:24
  and
• 11:24
  more.
• 11:25
  The
• 11:25
  BMAD
• 11:26
  Builder
• 11:26
  lets
• 11:27
  users
• 11:27
  create
• 11:27
  custom
• 11:28
  agents
• 11:28
  with
• 11:29
  memory
• 11:29
  and
• 11:29
  structured
• 11:30
  workflows.
• 11:31
  Meanwhile,
• 11:32
  MIT's
• 11:32
  Missing
• 11:32
  Semester
• 11:33
  course
• 11:33
  returns
• 11:33
  for
• 11:34
  2026
• 11:35
  with
• 11:35
  AI
• 11:35
  tools
• 11:36
  integrated
• 11:36
  directly
• 11:37
  into
• 11:37
  every
• 11:37
  lecture
• 11:38
  rather
• 11:38
  than
• 11:38
  siloed
• 11:39
  into
• 11:39
  a
• 11:39
  standalone
• 11:40
  module
• 11:40
  --
• 11:41
  acknowledging
• 11:41
  that
• 11:41
  command-line
• 11:42
  proficiency,
• 11:43
  version
• 11:43
  control,
• 11:44
  and
• 11:44
  debugging
• 11:45
  are
• 11:45
  now
• 11:45
  inseparable
• 11:46
  from
• 11:46
  AI-assisted
• 11:47
  workflows.
• 11:50
  The
• 11:50
  gap
• 11:50
  between
• 11:51
  "I
• 11:51
  have
• 11:51
  an
• 11:51
  agent"
• 11:52
  and
• 11:52
  "I
• 11:53
  operate
• 11:53
  a
• 11:53
  fleet
• 11:53
  of
• 11:54
  agents"
• 11:54
  is
• 11:55
  spawning
• 11:55
  a
• 11:55
  new
• 11:55
  category
• 11:56
  of
• 11:56
  infrastructure
• 11:56
  tooling.
• 11:57
  Klaw
• 11:57
  bills
• 11:58
  itself
• 11:58
  as
• 11:58
  "kubectl
• 11:59
  for
• 11:59
  AI
• 11:59
  agents"
• 11:59
  --
• 12:00
  a
• 12:00
  single
• 12:01
  Go
• 12:01
  binary
• 12:01
  that
• 12:02
  brings
• 12:02
  Kubernetes-style
• 12:03
  operations
• 12:03
  to
• 12:04
  AI
• 12:04
  agent
• 12:04
  management.
• 12:05
  Agents
• 12:05
  are
• 12:06
  organized
• 12:06
  into
• 12:06
  namespaces
• 12:07
  with
• 12:07
  scoped
• 12:07
  secrets
• 12:08
  and
• 12:08
  tool
• 12:08
  permissions
• 12:08
  (sales
• 12:09
  agents
• 12:09
  can't
• 12:09
  access
• 12:10
  support
• 12:10
  API
• 12:10
  keys),
• 12:11
  managed
• 12:11
  via
• 12:12
  familiar
• 12:12
  commands
• 12:12
  like
• 12:12
  `klaw
• 12:13
  get
• 12:13
  agents`,
• 12:13
  `klaw
• 12:14
  describe
• 12:14
  agent`,
• 12:15
  and
• 12:15
  `klaw
• 12:15
  logs
• 12:15
  --follow`.
• 12:16
  Built-in
• 12:17
  cron
• 12:17
  scheduling
• 12:17
  replaces
• 12:18
  messy
• 12:18
  Lambda
• 12:18
  functions.
• 12:19
  Slack
• 12:20
  integration
• 12:20
  turns
• 12:20
  a
• 12:21
  workspace
• 12:21
  into
• 12:21
  a
• 12:21
  command
• 12:22
  center
• 12:22
  where
• 12:22
  non-technical
• 12:23
  stakeholders
• 12:23
  can
• 12:23
  query
• 12:24
  agent
• 12:24
  status,
• 12:25
  trigger
• 12:25
  runs,
• 12:26
  and
• 12:26
  issue
• 12:26
  ad-hoc
• 12:27
  tasks
• 12:27
  via
• 12:27
  `@klaw`
• 12:28
  mentions.
• 12:29
  Distributed
• 12:30
  mode
• 12:30
  supports
• 12:30
  a
• 12:30
  central
• 12:30
  controller
• 12:31
  with
• 12:31
  worker
• 12:31
  nodes
• 12:32
  joining
• 12:32
  via
• 12:32
  `klaw
• 12:32
  node
• 12:33
  join`.
• 12:33
  The
• 12:33
  important
• 12:34
  caveat:
• 12:35
  namespaces
• 12:35
  provide
• 12:36
  logical
• 12:36
  isolation
• 12:37
  only
• 12:37
  --
• 12:38
  agents
• 12:38
  run
• 12:38
  under
• 12:39
  the
• 12:39
  host
• 12:39
  user
• 12:39
  account
• 12:39
  without
• 12:40
  filesystem
• 12:40
  sandboxing
• 12:41
  unless
• 12:41
  explicitly
• 12:42
  containerized.
• 12:43
  TokenHub
• 12:43
  addresses
• 12:44
  a
• 12:44
  complementary
• 12:44
  problem:
• 12:45
  routing
• 12:46
  requests
• 12:46
  across
• 12:47
  multiple
• 12:47
  LLM
• 12:47
  providers
• 12:48
  to
• 12:48
  optimize
• 12:48
  cost,
• 12:49
  latency,
• 12:49
  reliability,
• 12:50
  and
• 12:50
  quality.
• 12:51
  Its
• 12:51
  multi-objective
• 12:52
  routing
• 12:52
  engine
• 12:53
  filters
• 12:53
  eligible
• 12:54
  models
• 12:54
  by
• 12:54
  budget,
• 12:55
  latency,
• 12:55
  context
• 12:56
  window,
• 12:56
  and
• 12:56
  health,
• 12:57
  then
• 12:57
  scores
• 12:57
  candidates
• 12:58
  using
• 12:58
  weighted
• 12:58
  profiles
• 12:59
  (cheap,
• 12:59
  normal,
• 13:00
  high
• 13:00
  confidence,
• 13:01
  planning,
• 13:01
  adversarial).
• 13:03
  For
• 13:03
  adaptive
• 13:03
  selection,
• 13:04
  Thompson
• 13:04
  Sampling
• 13:05
  replaces
• 13:05
  deterministic
• 13:06
  scoring
• 13:06
  with
• 13:07
  probabilistic
• 13:07
  Beta-distribution
• 13:08
  sampling
• 13:08
  that
• 13:09
  learns
• 13:09
  from
• 13:09
  reward
• 13:09
  logs.
• 13:10
  Three
• 13:10
  orchestration
• 13:11
  modes
• 13:11
  --
• 13:12
  adversarial
• 13:13
  (plan/critique/refine),
• 13:15
  vote
• 13:15
  (fan-out
• 13:16
  to
• 13:16
  N
• 13:16
  models
• 13:16
  with
• 13:16
  a
• 13:16
  judge),
• 13:17
  and
• 13:17
  refine
• 13:18
  (iterative
• 13:18
  self-improvement)
• 13:18
  --
• 13:19
  support
• 13:20
  complex
• 13:20
  multi-model
• 13:21
  workflows.
• 13:22
  Security
• 13:23
  includes
• 13:23
  an
• 13:23
  AES-256-GCM
• 13:27
  encrypted
• 13:27
  credential
• 13:28
  vault
• 13:28
  with
• 13:28
  Argon2id
• 13:29
  key
• 13:29
  derivation
• 13:30
  and
• 13:30
  per-key
• 13:31
  monthly
• 13:31
  budget
• 13:31
  enforcement.
• 13:32
  The
• 13:33
  admin
• 13:33
  UI
• 13:33
  features
• 13:34
  a
• 13:34
  real-time
• 13:34
  Cytoscape.js
• 13:38
  flow
• 13:38
  graph
• 13:38
  and
• 13:38
  a
• 13:38
  what-if
• 13:39
  routing
• 13:39
  simulator
• 13:40
  for
• 13:40
  testing
• 13:40
  model
• 13:40
  selection
• 13:41
  without
• 13:41
  live
• 13:41
  requests.
• 13:42
  Reuven
• 13:43
  Cohen
• 13:43
  argues
• 13:43
  the
• 13:43
  real
• 13:44
  agent
• 13:44
  stack
• 13:44
  starts
• 13:45
  in
• 13:45
  your
• 13:45
  pocket.
• 13:46
  The
• 13:46
  phone
• 13:46
  is
• 13:46
  the
• 13:46
  ultimate
• 13:47
  edge
• 13:47
  device
• 13:47
  --
• 13:48
  sensors,
• 13:49
  biometrics,
• 13:49
  camera,
• 13:50
  microphone,
• 13:50
  secure
• 13:51
  enclave,
• 13:51
  payment
• 13:52
  rails.
• 13:53
  Mobile-first
• 13:53
  agents
• 13:54
  aren't
• 13:54
  chatbots
• 13:54
  in
• 13:54
  an
• 13:55
  app;
• 13:55
  they're
• 13:55
  system-level
• 13:56
  actors
• 13:57
  watching
• 13:57
  notifications,
• 13:58
  calendar
• 13:58
  events,
• 13:59
  and
• 13:59
  intent
• 13:59
  signals,
• 14:00
  running
• 14:00
  local
• 14:01
  reasoning
• 14:01
  loops
• 14:02
  and
• 14:02
  escalating
• 14:02
  to
• 14:03
  cloud
• 14:03
  only
• 14:03
  when
• 14:04
  the
• 14:04
  task
• 14:04
  demands
• 14:04
  scale.
• 14:05
  His
• 14:06
  scenario:
• 14:07
  a
• 14:07
  field
• 14:07
  technician
• 14:08
  carrying
• 14:08
  a
• 14:08
  foldable
• 14:09
  phone
• 14:09
  running
• 14:09
  agents
• 14:10
  that
• 14:10
  store
• 14:10
  repair
• 14:11
  embeddings
• 14:11
  locally
• 14:12
  for
• 14:12
  similarity
• 14:12
  search,
• 14:13
  diagnose
• 14:14
  equipment
• 14:14
  from
• 14:14
  camera
• 14:15
  input,
• 14:15
  and
• 14:15
  sync
• 14:16
  verified
• 14:16
  updates
• 14:17
  when
• 14:17
  connectivity
• 14:17
  returns.
• 14:19
  No
• 14:19
  cloud
• 14:19
  required.
• 14:20
  Mobile-first
• 14:21
  forces
• 14:21
  discipline
• 14:21
  --
• 14:23
  power
• 14:23
  limits,
• 14:24
  latency
• 14:24
  ceilings,
• 14:25
  privacy
• 14:25
  constraints
• 14:26
  --
• 14:27
  that
• 14:27
  produces
• 14:27
  better
• 14:27
  architecture.
• 14:31
  Liquid
• 14:31
  AI
• 14:31
  continues
• 14:32
  scaling
• 14:32
  its
• 14:33
  non-transformer
• 14:33
  architecture
• 14:34
  with
• 14:34
  LFM2-24B-A2B,
• 14:38
  a
• 14:38
  sparse
• 14:38
  Mixture-of-Experts
• 14:39
  model
• 14:39
  with
• 14:39
  24
• 14:40
  billion
• 14:40
  total
• 14:40
  parameters
• 14:41
  but
• 14:41
  only
• 14:42
  2.3
• 14:42
  billion
• 14:43
  active
• 14:43
  per
• 14:43
  forward
• 14:44
  pass.
• 14:45
  The
• 14:45
  architecture
• 14:45
  stacks
• 14:46
  40
• 14:46
  layers
• 14:46
  with
• 14:46
  64
• 14:47
  experts
• 14:47
  per
• 14:48
  MoE
• 14:48
  block
• 14:48
  using
• 14:49
  top-4
• 14:49
  routing,
• 14:50
  maintaining
• 14:50
  the
• 14:50
  hybrid
• 14:51
  convolution
• 14:51
  plus
• 14:51
  grouped-query
• 14:52
  attention
• 14:52
  design
• 14:53
  that
• 14:53
  defined
• 14:53
  the
• 14:53
  smaller
• 14:54
  LFM2
• 14:54
  variants.
• 14:55
  Across
• 14:56
  benchmarks
• 14:56
  including
• 14:57
  GPQA
• 14:57
  Diamond,
• 14:58
  MMLU-Pro,
• 15:00
  IFEval,
• 15:00
  and
• 15:00
  MATH-500,
• 15:02
  quality
• 15:03
  improves
• 15:03
  log-linearly
• 15:04
  from
• 15:04
  350M
• 15:05
  to
• 15:05
  24B,
• 15:06
  confirming
• 15:06
  the
• 15:07
  architecture
• 15:07
  doesn't
• 15:07
  plateau
• 15:08
  at
• 15:08
  small
• 15:08
  sizes.
• 15:09
  Designed
• 15:10
  to
• 15:10
  run
• 15:10
  within
• 15:10
  32GB
• 15:11
  RAM
• 15:11
  with
• 15:11
  day-zero
• 15:12
  support
• 15:12
  for
• 15:12
  llama.cpp,
• 15:14
  vLLM,
• 15:15
  and
• 15:15
  SGLang,
• 15:16
  this
• 15:16
  is
• 15:16
  the
• 15:16
  scale
• 15:16
  test
• 15:17
  for
• 15:17
  Liquid's
• 15:17
  thesis
• 15:17
  that
• 15:18
  concentrating
• 15:18
  capacity
• 15:19
  in
• 15:19
  total
• 15:19
  parameters
• 15:20
  rather
• 15:20
  than
• 15:20
  active
• 15:21
  compute
• 15:21
  can
• 15:21
  deliver
• 15:21
  competitive
• 15:22
  quality
• 15:22
  while
• 15:22
  keeping
• 15:23
  inference
• 15:23
  efficient
• 15:23
  enough
• 15:24
  for
• 15:24
  high-end
• 15:24
  consumer
• 15:24
  hardware.
• 15:26
  Qwen3's
• 15:26
  voice
• 15:26
  capabilities
• 15:27
  keep
• 15:27
  gaining
• 15:27
  depth.
• 15:28
  The
• 15:29
  TTS
• 15:29
  model
• 15:29
  uses
• 15:30
  voice
• 15:30
  embeddings
• 15:31
  --
• 15:32
  1024-dimensional
• 15:32
  vectors
• 15:33
  (2048
• 15:33
  for
• 15:34
  the
• 15:34
  1.7B
• 15:34
  variant)
• 15:34
  --
• 15:36
  that
• 15:36
  enable
• 15:36
  voice
• 15:36
  cloning,
• 15:37
  but
• 15:37
  the
• 15:37
  more
• 15:37
  interesting
• 15:38
  application
• 15:38
  is
• 15:39
  vector
• 15:39
  arithmetic
• 15:39
  on
• 15:40
  voices.
• 15:41
  Gender
• 15:41
  swapping,
• 15:42
  pitch
• 15:42
  modification,
• 15:43
  voice
• 15:43
  mixing,
• 15:44
  emotion
• 15:44
  space
• 15:44
  creation,
• 15:45
  and
• 15:45
  semantic
• 15:46
  voice
• 15:46
  search
• 15:46
  all
• 15:46
  become
• 15:47
  possible
• 15:47
  through
• 15:47
  standard
• 15:48
  vector
• 15:48
  operations.
• 15:49
  The
• 15:49
  voice
• 15:49
  embedding
• 15:50
  model
• 15:50
  is
• 15:50
  a
• 15:50
  tiny
• 15:51
  encoder
• 15:51
  with
• 15:51
  just
• 15:51
  a
• 15:51
  few
• 15:52
  million
• 15:52
  parameters,
• 15:53
  now
• 15:53
  ripped
• 15:53
  out
• 15:54
  as
• 15:54
  a
• 15:54
  standalone
• 15:54
  model
• 15:55
  with
• 15:55
  ONNX
• 15:56
  exports
• 15:56
  for
• 15:56
  optimized
• 15:57
  web
• 15:57
  inference.
• 15:58
  The
• 15:58
  Crane
• 15:58
  inference
• 15:59
  engine
• 15:59
  makes
• 15:59
  Qwen3-TTS
• 16:00
  accessible
• 16:01
  from
• 16:01
  a
• 16:01
  pure
• 16:01
  Rust
• 16:01
  codebase
• 16:02
  --
• 16:02
  no
• 16:03
  C++
• 16:03
  glue,
• 16:04
  no
• 16:04
  Python
• 16:04
  runtime,
• 16:05
  no
• 16:05
  GGUF
• 16:06
  conversion.
• 16:07
  Built
• 16:07
  on
• 16:07
  Candle,
• 16:08
  it
• 16:08
  supports
• 16:08
  CPU,
• 16:09
  CUDA,
• 16:09
  and
• 16:10
  Metal
• 16:10
  with
• 16:10
  up
• 16:10
  to
• 16:10
  6x
• 16:11
  faster
• 16:11
  inference
• 16:11
  than
• 16:12
  vanilla
• 16:12
  PyTorch
• 16:12
  on
• 16:12
  Apple
• 16:13
  Silicon.
• 16:14
  The
• 16:14
  supported
• 16:14
  model
• 16:15
  list
• 16:15
  now
• 16:15
  includes
• 16:15
  Qwen
• 16:16
  2.5/3,
• 16:17
  Hunyuan,
• 16:18
  Qwen-VL,
• 16:18
  PaddleOCR,
• 16:19
  Moonshine
• 16:20
  ASR,
• 16:20
  Silero
• 16:21
  VAD,
• 16:21
  and
• 16:22
  Qwen3-TTS
• 16:23
  with
• 16:23
  a
• 16:23
  native
• 16:23
  speech-tokenizer
• 16:24
  decoder.
• 16:25
  A
• 16:25
  sobering
• 16:26
  finding
• 16:26
  from
• 16:26
  on-device
• 16:26
  testing:
• 16:27
  the
• 16:28
  same
• 16:28
  INT8
• 16:28
  ONNX
• 16:29
  model
• 16:29
  deployed
• 16:30
  to
• 16:30
  five
• 16:30
  different
• 16:30
  Snapdragon
• 16:31
  chipsets
• 16:31
  produced
• 16:32
  accuracy
• 16:32
  ranging
• 16:33
  from
• 16:33
  93%
• 16:34
  on
• 16:34
  the
• 16:34
  8
• 16:34
  Gen
• 16:34
  3
• 16:35
  down
• 16:35
  to
• 16:35
  71%
• 16:36
  on
• 16:36
  the
• 16:36
  4
• 16:36
  Gen
• 16:36
  2,
• 16:37
  against
• 16:37
  a
• 16:38
  cloud
• 16:38
  benchmark
• 16:38
  of
• 16:38
  94.2%.
• 16:41
  The
• 16:41
  spread
• 16:41
  comes
• 16:41
  from
• 16:42
  three
• 16:42
  factors:
• 16:43
  INT8
• 16:44
  rounding
• 16:44
  behavior
• 16:45
  differs
• 16:45
  across
• 16:46
  Hexagon
• 16:46
  NPU
• 16:47
  generations
• 16:47
  (not
• 16:48
  all
• 16:48
  INT8
• 16:49
  is
• 16:49
  created
• 16:49
  equal),
• 16:50
  the
• 16:50
  QNN
• 16:51
  runtime
• 16:51
  optimizes
• 16:52
  graphs
• 16:52
  differently
• 16:53
  per
• 16:53
  SoC
• 16:53
  trading
• 16:54
  accuracy
• 16:54
  for
• 16:54
  throughput,
• 16:55
  and
• 16:55
  on
• 16:55
  lower-tier
• 16:56
  chips
• 16:56
  certain
• 16:57
  ops
• 16:57
  fall
• 16:57
  back
• 16:57
  from
• 16:57
  NPU
• 16:58
  to
• 16:58
  CPU,
• 16:59
  changing
• 16:59
  the
• 17:00
  execution
• 17:00
  path
• 17:00
  entirely.
• 17:02
  None
• 17:02
  of
• 17:02
  this
• 17:02
  shows
• 17:03
  up
• 17:03
  in
• 17:03
  cloud-based
• 17:03
  benchmarks.
• 17:04
  Most
• 17:05
  CI
• 17:05
  pipelines
• 17:06
  test
• 17:06
  on
• 17:06
  cloud
• 17:06
  GPUs
• 17:07
  and
• 17:07
  call
• 17:08
  it
• 17:08
  a
• 17:08
  day
• 17:08
  --
• 17:09
  a
• 17:09
  blind
• 17:09
  spot
• 17:09
  that
• 17:10
  will
• 17:10
  only
• 17:10
  widen
• 17:10
  as
• 17:11
  on-device
• 17:11
  deployment
• 17:12
  accelerates.
• 17:15
  A
• 17:15
  new
• 17:15
  survey
• 17:15
  from
• 17:15
  Shandong
• 17:16
  University
• 17:16
  poses
• 17:17
  the
• 17:17
  question
• 17:17
  directly:
• 17:17
  what
• 17:18
  actually
• 17:18
  breaks
• 17:18
  embodied
• 17:19
  AI
• 17:19
  security
• 17:19
  --
• 17:20
  LLM
• 17:21
  vulnerabilities,
• 17:22
  cyber-physical
• 17:22
  system
• 17:22
  flaws,
• 17:23
  or
• 17:23
  something
• 17:23
  fundamentally
• 17:24
  different?
• 17:25
  The
• 17:25
  answer
• 17:25
  is
• 17:25
  the
• 17:26
  third
• 17:26
  option.
• 17:27
  The
• 17:27
  paper
• 17:27
  argues
• 17:28
  that
• 17:28
  treating
• 17:28
  embodied
• 17:29
  AI
• 17:29
  as
• 17:29
  either
• 17:29
  a
• 17:29
  "vulnerable
• 17:30
  brain"
• 17:31
  (LLM-centric
• 17:32
  threats)
• 17:32
  or
• 17:32
  a
• 17:32
  "fragile
• 17:32
  body"
• 17:33
  (classical
• 17:33
  CPS
• 17:34
  faults)
• 17:34
  misses
• 17:35
  a
• 17:35
  critical
• 17:35
  class
• 17:35
  of
• 17:35
  failures
• 17:36
  arising
• 17:36
  from
• 17:36
  embodiment-induced
• 17:37
  system-level
• 17:37
  mismatches.
• 17:39
  Four
• 17:39
  insights
• 17:39
  drive
• 17:40
  the
• 17:40
  argument:
• 17:41
  semantic
• 17:41
  correctness
• 17:41
  does
• 17:41
  not
• 17:42
  imply
• 17:42
  physical
• 17:42
  safety,
• 17:43
  because
• 17:43
  language
• 17:43
  abstracts
• 17:44
  away
• 17:44
  forces,
• 17:45
  friction,
• 17:45
  and
• 17:45
  timing;
• 17:46
  identical
• 17:46
  actions
• 17:47
  can
• 17:47
  produce
• 17:47
  drastically
• 17:48
  different
• 17:48
  consequences
• 17:48
  depending
• 17:49
  on
• 17:49
  physical
• 17:49
  state
• 17:50
  due
• 17:50
  to
• 17:50
  nonlinear
• 17:50
  dynamics;
• 17:51
  minor
• 17:52
  perception
• 17:52
  errors
• 17:52
  cascade
• 17:53
  across
• 17:53
  tightly
• 17:53
  coupled
• 17:54
  perception-decision-action
• 17:55
  loops,
• 17:55
  compounding
• 17:56
  gradually
• 17:56
  rather
• 17:56
  than
• 17:56
  failing
• 17:57
  catastrophically;
• 17:58
  and
• 17:58
  safety
• 17:59
  is
• 17:59
  not
• 17:59
  compositional
• 17:59
  --
• 18:00
  locally
• 18:01
  safe
• 18:01
  decisions
• 18:01
  can
• 18:01
  accumulate
• 18:02
  into
• 18:02
  globally
• 18:03
  unsafe
• 18:03
  trajectories.
• 18:04
  The
• 18:04
  taxonomy
• 18:05
  maps
• 18:05
  attacks
• 18:06
  from
• 18:06
  adversarial
• 18:06
  patches
• 18:07
  and
• 18:07
  visual
• 18:07
  command
• 18:07
  hijacking
• 18:08
  to
• 18:08
  environmental
• 18:09
  jailbreaking,
• 18:10
  where
• 18:10
  the
• 18:10
  physical
• 18:10
  surroundings
• 18:11
  themselves
• 18:11
  become
• 18:11
  the
• 18:12
  injection
• 18:12
  vector.
• 18:13
  A
• 18:13
  real-world
• 18:14
  example
• 18:14
  landed
• 18:14
  the
• 18:14
  same
• 18:15
  week.
• 18:16
  Software
• 18:16
  engineer
• 18:17
  Sammy
• 18:17
  Azdoufal
• 18:17
  was
• 18:17
  building
• 18:18
  a
• 18:18
  DIY
• 18:18
  controller
• 18:19
  for
• 18:19
  his
• 18:19
  DJI
• 18:19
  Romo
• 18:20
  robot
• 18:20
  vacuum
• 18:21
  when
• 18:21
  he
• 18:21
  discovered
• 18:21
  the
• 18:21
  backend
• 18:22
  credentials
• 18:22
  that
• 18:23
  authenticated
• 18:23
  his
• 18:23
  device
• 18:24
  also
• 18:24
  granted
• 18:24
  access
• 18:25
  to
• 18:25
  live
• 18:25
  camera
• 18:25
  feeds,
• 18:25
  microphone
• 18:26
  audio,
• 18:26
  floor
• 18:26
  maps,
• 18:27
  and
• 18:27
  status
• 18:27
  data
• 18:28
  from
• 18:28
  nearly
• 18:28
  7,000
• 18:29
  other
• 18:29
  vacuums
• 18:29
  across
• 18:29
  24
• 18:30
  countries.
• 18:30
  He
• 18:30
  wasn't
• 18:31
  hacking
• 18:31
  --
• 18:31
  just
• 18:32
  reverse-engineering
• 18:32
  the
• 18:33
  MQTT
• 18:33
  communication
• 18:34
  protocol
• 18:34
  with
• 18:34
  help
• 18:34
  from
• 18:34
  an
• 18:34
  AI
• 18:35
  coding
• 18:35
  assistant.
• 18:36
  The
• 18:36
  server
• 18:36
  simply
• 18:37
  failed
• 18:37
  to
• 18:37
  scope
• 18:37
  authentication
• 18:38
  tokens
• 18:38
  to
• 18:38
  individual
• 18:39
  devices,
• 18:39
  treating
• 18:40
  any
• 18:40
  valid
• 18:40
  token
• 18:40
  as
• 18:41
  authorization
• 18:41
  for
• 18:41
  the
• 18:41
  entire
• 18:42
  fleet.
• 18:43
  DJI
• 18:43
  says
• 18:43
  the
• 18:43
  issue
• 18:44
  was
• 18:44
  patched
• 18:44
  in
• 18:44
  two
• 18:45
  updates
• 18:45
  (February
• 18:45
  8
• 18:45
  and
• 18:46
  10),
• 18:46
  but
• 18:46
  the
• 18:47
  episode
• 18:47
  highlights
• 18:47
  the
• 18:47
  survey's
• 18:48
  Insight
• 18:48
  3
• 18:48
  in
• 18:48
  consumer
• 18:49
  hardware:
• 18:50
  a
• 18:50
  single
• 18:50
  authentication
• 18:51
  error
• 18:51
  propagated
• 18:51
  across
• 18:52
  the
• 18:52
  entire
• 18:52
  perception-action
• 18:53
  surface
• 18:53
  of
• 18:54
  thousands
• 18:54
  of
• 18:54
  internet-connected
• 18:55
  robots
• 18:55
  equipped
• 18:56
  with
• 18:56
  cameras,
• 18:57
  microphones,
• 18:57
  and
• 18:58
  detailed
• 18:58
  maps
• 18:58
  of
• 18:58
  private
• 18:59
  homes.
• 18:59
  The
• 19:00
  other
• 19:00
  side
• 19:00
  of
• 19:00
  this
• 19:00
  coin
• 19:01
  is
• 19:01
  accessibility.
• 19:02
  Bilawal
• 19:03
  Sidhu,
• 19:03
  a
• 19:03
  former
• 19:04
  Google
• 19:04
  PM
• 19:05
  who
• 19:05
  helped
• 19:05
  build
• 19:05
  Google
• 19:06
  Earth's
• 19:06
  photogrammetry
• 19:06
  pipeline,
• 19:07
  built
• 19:07
  WorldView
• 19:08
  --
• 19:09
  a
• 19:09
  browser-based
• 19:10
  satellite
• 19:10
  simulator
• 19:10
  layering
• 19:11
  real-time
• 19:11
  data
• 19:12
  feeds
• 19:12
  (OpenSky
• 19:12
  Network
• 19:13
  aircraft
• 19:13
  positions,
• 19:14
  CelesTrak
• 19:14
  satellite
• 19:15
  orbits,
• 19:16
  OpenStreetMap
• 19:16
  vehicle
• 19:17
  flow,
• 19:17
  and
• 19:17
  public
• 19:18
  CCTV
• 19:18
  cameras)
• 19:19
  over
• 19:19
  Google's
• 19:19
  volumetric
• 19:20
  3D
• 19:20
  city
• 19:21
  models.
• 19:22
  Night
• 19:22
  vision,
• 19:23
  FLIR
• 19:23
  thermal,
• 19:24
  CRT
• 19:24
  scan
• 19:25
  lines
• 19:25
  --
• 19:26
  all
• 19:26
  built
• 19:26
  from
• 19:26
  actual
• 19:27
  military
• 19:27
  display
• 19:27
  specifications.
• 19:29
  He
• 19:29
  built
• 19:29
  it
• 19:29
  in
• 19:29
  a
• 19:29
  weekend
• 19:30
  using
• 19:30
  multiple
• 19:31
  AI
• 19:31
  agents
• 19:31
  running
• 19:32
  simultaneously
• 19:32
  (Gemini
• 19:33
  3.1,
• 19:34
  Claude
• 19:34
  4.6,
• 19:35
  Codex
• 19:36
  5.3),
• 19:37
  sometimes
• 19:37
  eight
• 19:37
  at
• 19:38
  once
• 19:38
  working
• 19:38
  on
• 19:38
  different
• 19:38
  subsystems.
• 19:40
  The
• 19:40
  demo
• 19:40
  drew
• 19:41
  a
• 19:41
  response
• 19:41
  from
• 19:41
  Palantir
• 19:42
  co-founder
• 19:42
  Joe
• 19:42
  Lonsdale,
• 19:43
  and
• 19:43
  for
• 19:43
  good
• 19:44
  reason:
• 19:44
  it
• 19:45
  demonstrates
• 19:45
  that
• 19:45
  the
• 19:45
  visual
• 19:46
  language
• 19:46
  of
• 19:46
  classified
• 19:47
  intelligence
• 19:47
  systems
• 19:48
  can
• 19:48
  now
• 19:48
  run
• 19:48
  in
• 19:48
  a
• 19:48
  browser
• 19:48
  tab,
• 19:49
  built
• 19:49
  by
• 19:49
  one
• 19:50
  person
• 19:50
  in
• 19:50
  days.
• 19:51
  What
• 19:51
  used
• 19:51
  to
• 19:51
  be
• 19:51
  the
• 19:52
  exclusive
• 19:52
  infrastructure
• 19:53
  of
• 19:53
  intelligence
• 19:53
  agencies
• 19:54
  is
• 19:54
  becoming
• 19:55
  commoditized
• 19:55
  through
• 19:55
  the
• 19:55
  convergence
• 19:56
  of
• 19:56
  public
• 19:56
  data,
• 19:57
  3D
• 19:58
  mapping
• 19:58
  APIs,
• 19:59
  and
• 19:59
  AI-assisted
• 20:00
  development.