Latest / AGI Dreams – Open, Uncensored, & Local - AI News Digest / AGI Dreams Podcast – October 28, 2026

Transcript

AGI Dreams Podcast – October 28, 2026 · AGI Dreams – Open, Uncensored, & Local - AI News Digest

• 0:00
  Anthropic
• 0:00
  has
• 0:00
  published
• 0:01
  evidence
• 0:01
  that
• 0:01
  a
• 0:01
  frontier
• 0:02
  model
• 0:02
  can
• 0:02
  do
• 0:02
  the
• 0:03
  whole
• 0:03
  job,
• 0:03
  not
• 0:04
  just
• 0:04
  the
• 0:04
  first
• 0:04
  half
• 0:04
  of
• 0:05
  it.
• 0:05
  Using
• 0:06
  an
• 0:06
  unreleased
• 0:06
  internal
• 0:07
  research
• 0:07
  model,
• 0:08
  the
• 0:08
  company
• 0:08
  built
• 0:09
  five
• 0:09
  working
• 0:09
  attack
• 0:10
  chains
• 0:10
  against
• 0:10
  some
• 0:11
  of
• 0:11
  the
• 0:11
  most
• 0:11
  scrutinized
• 0:12
  software
• 0:12
  on
• 0:12
  the
• 0:12
  planet
• 0:13
  and
• 0:13
  reported
• 0:14
  every
• 0:14
  bug
• 0:15
  under
• 0:15
  a
• 0:15
  90-day
• 0:16
  disclosure
• 0:16
  window.
• 0:17
  The
• 0:18
  Chrome
• 0:18
  chain
• 0:18
  alone
• 0:19
  produced
• 0:19
  62
• 0:20
  CVEs,
• 0:21
  all
• 0:21
  rated
• 0:21
  high
• 0:21
  or
• 0:22
  critical,
• 0:22
  patched
• 0:23
  as
• 0:23
  of
• 0:23
  Chrome
• 0:23
  156,
• 0:25
  including
• 0:26
  13
• 0:26
  sandbox
• 0:27
  escapes;
• 0:28
  the
• 0:28
  proof-of-concept
• 0:29
  launches
• 0:29
  Calculator
• 0:30
  from
• 0:30
  a
• 0:30
  stock
• 0:30
  browser
• 0:30
  within
• 0:31
  seconds
• 0:31
  of
• 0:31
  a
• 0:31
  page
• 0:32
  load,
• 0:32
  with
• 0:32
  no
• 0:32
  further
• 0:33
  clicks.
• 0:33
  The
• 0:34
  Safari
• 0:34
  chain
• 0:34
  is
• 0:34
  more
• 0:35
  interesting
• 0:35
  technically
• 0:36
  because
• 0:36
  it
• 0:36
  never
• 0:36
  executes
• 0:37
  native
• 0:37
  code:
• 0:38
  it
• 0:38
  is
• 0:39
  a
• 0:39
  data-only
• 0:39
  attack
• 0:40
  that
• 0:40
  rewrites
• 0:40
  a
• 0:40
  security
• 0:41
  flag
• 0:41
  in
• 0:41
  memory
• 0:41
  to
• 0:42
  disable
• 0:42
  the
• 0:42
  same-origin
• 0:43
  policy,
• 0:44
  turning
• 0:44
  several
• 0:44
  intended
• 0:45
  behaviors
• 0:45
  into
• 0:45
  an
• 0:46
  exploit
• 0:46
  when
• 0:46
  combined.
• 0:47
  There
• 0:48
  is
• 0:48
  also
• 0:48
  an
• 0:48
  Android
• 0:48
  lock-screen
• 0:49
  bypass
• 0:50
  tested
• 0:50
  on
• 0:50
  a
• 0:50
  physical
• 0:51
  Pixel
• 0:51
  9,
• 0:52
  a
• 0:52
  Windows
• 0:52
  kernel
• 0:53
  QUIC
• 0:53
  flaw
• 0:54
  that
• 0:54
  remotely
• 0:54
  crashes
• 0:55
  HTTP/3
• 0:58
  servers,
• 0:58
  and
• 0:59
  a
• 0:59
  zero-click
• 0:59
  root
• 0:59
  on
• 1:00
  Linux
• 1:00
  that
• 1:00
  takes
• 1:01
  48
• 1:01
  minutes
• 1:01
  to
• 1:01
  defeat
• 1:02
  kernel
• 1:02
  address
• 1:02
  randomization
• 1:03
  and
• 1:04
  another
• 1:04
  five
• 1:04
  to
• 1:04
  land
• 1:05
  a
• 1:05
  shell
• 1:05
  as
• 1:05
  root
• 1:05
  .
• 1:06
  The
• 1:06
  methodology
• 1:07
  is
• 1:07
  the
• 1:07
  transferable
• 1:08
  lesson
• 1:08
  for
• 1:08
  defenders.
• 1:09
  Anthropic
• 1:10
  ran
• 1:10
  many
• 1:10
  scanning
• 1:11
  agents
• 1:11
  in
• 1:11
  parallel,
• 1:12
  scored
• 1:12
  each
• 1:13
  by
• 1:13
  how
• 1:13
  many
• 1:13
  net-new
• 1:14
  bugs
• 1:14
  it
• 1:14
  contributed,
• 1:15
  and
• 1:15
  tracked
• 1:15
  coverage
• 1:16
  so
• 1:16
  the
• 1:16
  swarm
• 1:17
  chewed
• 1:17
  through
• 1:17
  the
• 1:17
  whole
• 1:17
  codebase,
• 1:18
  with
• 1:18
  automated
• 1:19
  variant
• 1:19
  analysis
• 1:20
  on
• 1:20
  every
• 1:21
  confirmed
• 1:21
  finding
• 1:22
  and
• 1:22
  a
• 1:22
  standardized
• 1:23
  verification
• 1:23
  pipeline.
• 1:24
  The
• 1:24
  second
• 1:25
  recommendation
• 1:25
  is
• 1:25
  to
• 1:26
  scan
• 1:26
  at
• 1:26
  pull-request
• 1:27
  time
• 1:27
  via
• 1:27
  a
• 1:27
  Claude
• 1:28
  Code
• 1:28
  plugin,
• 1:29
  because
• 1:29
  clearing
• 1:30
  the
• 1:30
  backlog
• 1:30
  is
• 1:30
  useless
• 1:31
  if
• 1:31
  new
• 1:31
  bugs
• 1:31
  enter
• 1:32
  faster
• 1:32
  than
• 1:32
  you
• 1:32
  fix
• 1:33
  them.
• 1:33
  The
• 1:33
  expanded
• 1:34
  cyber
• 1:34
  program
• 1:35
  now
• 1:35
  exposes
• 1:35
  Claude
• 1:36
  Opus
• 1:36
  5.5,
• 1:37
  Sonnet
• 1:38
  5.5,
• 1:39
  and
• 1:39
  Mythos
• 1:39
  5.1
• 1:40
  with
• 1:40
  reduced
• 1:41
  blocking
• 1:41
  classifiers
• 1:42
  to
• 1:42
  qualifying
• 1:42
  security
• 1:43
  professionals.
• 1:44
  The
• 1:44
  capability
• 1:45
  line
• 1:45
  has
• 1:45
  moved.
• 1:46
  Earlier
• 1:47
  in
• 1:47
  the
• 1:47
  year
• 1:47
  the
• 1:47
  honest
• 1:48
  read
• 1:48
  was
• 1:48
  that
• 1:48
  these
• 1:48
  models
• 1:49
  found
• 1:49
  bugs
• 1:49
  brilliantly
• 1:50
  but
• 1:51
  mostly
• 1:51
  could
• 1:51
  not
• 1:51
  weaponize
• 1:52
  them
• 1:52
  past
• 1:52
  a
• 1:52
  sandbox
• 1:53
  on
• 1:53
  a
• 1:53
  few
• 1:53
  dollars
• 1:54
  of
• 1:54
  API
• 1:54
  credit.
• 1:55
  Full
• 1:55
  one-click
• 1:56
  chains
• 1:56
  in
• 1:56
  Chrome
• 1:57
  and
• 1:57
  Safari
• 1:57
  retire
• 1:58
  that
• 1:58
  caveat.
• 1:59
  The
• 1:59
  same
• 1:59
  shift
• 2:00
  explains
• 2:00
  the
• 2:00
  uproar
• 2:01
  over
• 2:01
  a
• 2:01
  $50,000
• 2:02
  PlayStation
• 2:03
  5
• 2:03
  bounty.
• 2:04
  A
• 2:04
  researcher
• 2:05
  known
• 2:05
  as
• 2:05
  Jordy
• 2:05
  used
• 2:06
  generative
• 2:06
  AI
• 2:06
  to
• 2:06
  rediscover
• 2:07
  a
• 2:07
  hypervisor
• 2:08
  bug
• 2:08
  affecting
• 2:08
  firmware
• 2:09
  13.60
• 2:10
  through
• 2:11
  14.00,
• 2:12
  reported
• 2:13
  it
• 2:13
  through
• 2:13
  HackerOne,
• 2:14
  and
• 2:14
  collected
• 2:15
  the
• 2:15
  payout,
• 2:15
  infuriating
• 2:16
  a
• 2:16
  homebrew
• 2:17
  scene
• 2:17
  that
• 2:17
  wanted
• 2:17
  it
• 2:17
  held
• 2:18
  until
• 2:18
  after
• 2:18
  GTA
• 2:19
  6
• 2:19
  shipped.
• 2:20
  Andy
• 2:20
  Nguyen,
• 2:21
  who
• 2:21
  found
• 2:21
  the
• 2:22
  same
• 2:22
  bug
• 2:22
  in
• 2:22
  2025,
• 2:23
  is
• 2:24
  quitting,
• 2:24
  calling
• 2:24
  the
• 2:25
  scene
• 2:25
  "a
• 2:25
  bunch
• 2:25
  of
• 2:25
  noobs
• 2:26
  using
• 2:26
  LLMs
• 2:27
  and
• 2:27
  writing
• 2:27
  hacks
• 2:27
  they
• 2:27
  don't
• 2:28
  even
• 2:28
  understand."
• 2:29
  Jordy's
• 2:30
  own
• 2:30
  logic
• 2:30
  is
• 2:30
  the
• 2:30
  uncomfortable
• 2:31
  truth
• 2:31
  of
• 2:31
  what
• 2:32
  observers
• 2:32
  are
• 2:32
  calling
• 2:32
  Black
• 2:33
  October:
• 2:34
  "If
• 2:34
  I
• 2:34
  and
• 2:34
  another
• 2:35
  person
• 2:35
  can
• 2:35
  find
• 2:35
  it
• 2:35
  with
• 2:36
  LLM,
• 2:36
  so
• 2:37
  can
• 2:37
  someone
• 2:37
  else"
• 2:37
  .
• 2:38
  The
• 2:38
  genre
• 2:39
  also
• 2:39
  breeds
• 2:39
  thin
• 2:40
  repackaging,
• 2:40
  like
• 2:41
  ARTEX,
• 2:41
  a
• 2:42
  Go-and-Next.js
• 2:43
  "autonomous
• 2:44
  penetration-testing
• 2:45
  system"
• 2:45
  trailing
• 2:45
  a
• 2:45
  challenge-winner
• 2:46
  claim
• 2:47
  but
• 2:47
  almost
• 2:47
  no
• 2:47
  documentation
• 2:48
  .
• 2:52
  The
• 2:52
  same
• 2:52
  capability
• 2:53
  runs
• 2:53
  in
• 2:54
  both
• 2:54
  directions,
• 2:55
  and
• 2:55
  the
• 2:55
  consumer-facing
• 2:56
  stories
• 2:56
  are
• 2:56
  messier
• 2:57
  than
• 2:57
  the
• 2:57
  lab's.
• 2:58
  A
• 2:58
  widely
• 2:58
  shared
• 2:59
  account
• 2:59
  describes
• 3:00
  two
• 3:00
  Macs
• 3:00
  compromised
• 3:01
  by
• 3:01
  a
• 3:01
  fake
• 3:01
  Chrome
• 3:02
  extension
• 3:02
  named
• 3:02
  "Notion
• 3:03
  Web
• 3:03
  Clipper,"
• 3:04
  a
• 3:04
  12-user
• 3:04
  impostor
• 3:05
  riding
• 3:05
  the
• 3:05
  name
• 3:06
  of
• 3:06
  a
• 3:06
  legitimate
• 3:06
  extension
• 3:07
  with
• 3:07
  a
• 3:07
  million
• 3:07
  users.
• 3:08
  Chrome
• 3:09
  Sync
• 3:09
  did
• 3:09
  the
• 3:09
  lateral
• 3:09
  movement
• 3:10
  for
• 3:10
  free:
• 3:11
  logging
• 3:11
  into
• 3:11
  the
• 3:11
  browser
• 3:12
  account
• 3:12
  installed
• 3:13
  the
• 3:13
  malicious
• 3:13
  extension
• 3:14
  on
• 3:14
  the
• 3:14
  second
• 3:14
  machine
• 3:15
  automatically.
• 3:16
  On
• 3:16
  the
• 3:16
  older
• 3:17
  M1
• 3:17
  Pro,
• 3:17
  private
• 3:18
  keys
• 3:18
  were
• 3:18
  pulled
• 3:19
  from
• 3:19
  a
• 3:19
  MetaMask
• 3:19
  wallet
• 3:20
  and
• 3:20
  roughly
• 3:20
  $230,000
• 3:22
  in
• 3:22
  Ethereum
• 3:22
  left
• 3:23
  at
• 3:23
  4
• 3:23
  AM;
• 3:24
  the
• 3:24
  M5
• 3:24
  Pro
• 3:25
  lost
• 3:25
  saved
• 3:25
  Safari
• 3:26
  and
• 3:26
  Chrome
• 3:26
  passwords
• 3:27
  plus
• 3:27
  2FA
• 3:28
  backup
• 3:28
  codes,
• 3:29
  with
• 3:29
  telemetry
• 3:30
  beaconing
• 3:30
  a
• 3:30
  location
• 3:31
  to
• 3:31
  a
• 3:31
  Tencent
• 3:31
  Cloud
• 3:32
  server
• 3:32
  in
• 3:32
  Shenzhen
• 3:33
  tied
• 3:33
  to
• 3:33
  the
• 3:33
  name
• 3:34
  "Li
• 3:34
  Wei."
• 3:35
  Claude
• 3:35
  Code
• 3:35
  identified
• 3:36
  the
• 3:36
  backdoor
• 3:36
  during
• 3:37
  a
• 3:37
  system
• 3:37
  audit,
• 3:38
  which
• 3:38
  is
• 3:38
  the
• 3:38
  part
• 3:38
  being
• 3:39
  celebrated
• 3:39
  .
• 3:40
  The
• 3:41
  sober
• 3:41
  framing
• 3:41
  is
• 3:41
  that
• 3:42
  the
• 3:42
  model
• 3:42
  found
• 3:42
  the
• 3:42
  implant
• 3:43
  after
• 3:43
  the
• 3:43
  money
• 3:43
  was
• 3:44
  gone,
• 3:44
  and
• 3:45
  finding
• 3:45
  one
• 3:45
  implant
• 3:46
  is
• 3:46
  not
• 3:46
  evidence
• 3:46
  the
• 3:46
  machines
• 3:47
  are
• 3:47
  clean.
• 3:49
  A
• 3:49
  browser
• 3:49
  extension
• 3:50
  spreading
• 3:50
  through
• 3:50
  an
• 3:50
  automatic
• 3:51
  sync
• 3:51
  mechanism
• 3:52
  is
• 3:52
  a
• 3:52
  vector
• 3:53
  the
• 3:53
  industry
• 3:53
  has
• 3:53
  watched
• 3:54
  drain
• 3:54
  wallets
• 3:54
  before,
• 3:55
  and
• 3:56
  the
• 3:56
  lesson
• 3:56
  is
• 3:56
  still
• 3:56
  about
• 3:57
  the
• 3:57
  update
• 3:57
  channel,
• 3:58
  not
• 3:58
  the
• 3:59
  model
• 3:59
  that
• 3:59
  cleaned
• 3:59
  up
• 3:59
  afterward.
• 4:01
  The
• 4:01
  offensive
• 4:01
  side
• 4:02
  of
• 4:02
  this
• 4:02
  living-room
• 4:02
  story
• 4:03
  is
• 4:03
  more
• 4:03
  striking.
• 4:05
  One
• 4:05
  viral
• 4:05
  post
• 4:06
  credits
• 4:06
  Claude
• 4:06
  with
• 4:06
  catching
• 4:07
  a
• 4:07
  North
• 4:07
  Korean
• 4:07
  supply-chain
• 4:08
  backdoor
• 4:09
  that
• 4:09
  "disguises
• 4:10
  RCE
• 4:10
  as
• 4:10
  an
• 4:10
  auth
• 4:11
  error,"
• 4:11
  hiding
• 4:12
  an
• 4:12
  eval
• 4:12
  call
• 4:12
  behind
• 4:13
  what
• 4:13
  looked
• 4:13
  like
• 4:13
  a
• 4:13
  routine
• 4:14
  failure
• 4:14
  path
• 4:14
  .
• 4:15
  On
• 4:15
  the
• 4:15
  other
• 4:16
  end
• 4:16
  of
• 4:16
  the
• 4:16
  spectrum
• 4:16
  sits
• 4:17
  a
• 4:17
  GitLab
• 4:17
  project
• 4:18
  published
• 4:18
  October
• 4:18
  24
• 4:19
  that
• 4:19
  packages
• 4:20
  an
• 4:20
  autonomous
• 4:20
  exploit
• 4:21
  chain
• 4:21
  for
• 4:21
  CVE-2022-22706
• 4:26
  against
• 4:26
  a
• 4:26
  TCL
• 4:27
  Android
• 4:27
  TV's
• 4:27
  Mali
• 4:28
  GPU,
• 4:29
  MIT-licensed
• 4:29
  and
• 4:30
  ready
• 4:30
  to
• 4:30
  run
• 4:30
  .
• 4:31
  A
• 4:31
  companion
• 4:32
  video
• 4:32
  shows
• 4:32
  Claude
• 4:33
  Code
• 4:33
  de-bloating
• 4:34
  Android
• 4:34
  and
• 4:34
  Google
• 4:34
  TVs
• 4:35
  from
• 4:35
  a
• 4:35
  single
• 4:35
  pasted
• 4:36
  prompt
• 4:36
  once
• 4:37
  the
• 4:37
  laptop
• 4:37
  and
• 4:37
  television
• 4:38
  share
• 4:38
  a
• 4:38
  Wi-Fi
• 4:39
  network
• 4:39
  .
• 4:40
  Re-exploiting
• 4:40
  a
• 4:40
  known
• 4:41
  2022
• 4:42
  vulnerability
• 4:42
  on
• 4:42
  a
• 4:42
  smart
• 4:43
  TV
• 4:43
  is
• 4:43
  the
• 4:43
  checklist
• 4:44
  case,
• 4:44
  not
• 4:45
  frontier
• 4:45
  research,
• 4:46
  but
• 4:46
  the
• 4:46
  tension
• 4:46
  is
• 4:47
  real:
• 4:47
  a
• 4:47
  model
• 4:48
  that
• 4:48
  refuses
• 4:48
  innocuous
• 4:49
  security
• 4:49
  questions
• 4:50
  as
• 4:50
  too
• 4:50
  dangerous
• 4:51
  will,
• 4:51
  in
• 4:51
  another
• 4:52
  session,
• 4:52
  cheerfully
• 4:53
  write
• 4:53
  a
• 4:53
  working
• 4:54
  GPU
• 4:54
  root
• 4:54
  chain
• 4:55
  for
• 4:55
  an
• 4:55
  appliance.
• 4:59
  The
• 4:59
  firmware
• 5:00
  story
• 5:00
  is
• 5:00
  the
• 5:00
  one
• 5:00
  that
• 5:01
  should
• 5:01
  keep
• 5:01
  infrastructure
• 5:02
  teams
• 5:02
  up
• 5:02
  at
• 5:03
  night.
• 5:03
  A
• 5:04
  buffer
• 5:04
  overflow
• 5:04
  rated
• 5:05
  9.5
• 5:06
  sits
• 5:06
  in
• 5:06
  TianoCore
• 5:07
  EDK2,
• 5:08
  the
• 5:08
  reference
• 5:09
  code
• 5:09
  behind
• 5:09
  UEFI
• 5:10
  firmware
• 5:10
  for
• 5:10
  essentially
• 5:11
  every
• 5:11
  server
• 5:12
  vendor,
• 5:12
  and
• 5:13
  it
• 5:13
  has
• 5:13
  been
• 5:13
  there
• 5:13
  for
• 5:13
  sixteen
• 5:14
  years.
• 5:15
  The
• 5:15
  flaw
• 5:15
  lives
• 5:15
  in
• 5:15
  the
• 5:15
  PXE
• 5:16
  Base
• 5:16
  Code
• 5:16
  driver
• 5:17
  that
• 5:17
  lets
• 5:17
  machines
• 5:18
  boot
• 5:18
  over
• 5:18
  the
• 5:18
  network:
• 5:19
  during
• 5:20
  PXE
• 5:20
  boot
• 5:20
  the
• 5:20
  driver
• 5:21
  copies
• 5:21
  each
• 5:21
  incoming
• 5:22
  ICMP
• 5:23
  error
• 5:23
  message
• 5:23
  into
• 5:24
  a
• 5:24
  fixed
• 5:24
  504-byte
• 5:25
  field,
• 5:26
  and
• 5:26
  the
• 5:26
  bounds
• 5:26
  handling
• 5:27
  is
• 5:27
  wrong
• 5:27
  in
• 5:27
  both
• 5:28
  the
• 5:28
  IPv4
• 5:29
  and
• 5:29
  IPv6
• 5:30
  handlers,
• 5:31
  advancing
• 5:31
  the
• 5:31
  destination
• 5:32
  pointer
• 5:32
  by
• 5:32
  cumulative
• 5:33
  rather
• 5:33
  than
• 5:33
  per-fragment
• 5:34
  length
• 5:34
  so
• 5:35
  an
• 5:35
  oversized
• 5:35
  message
• 5:36
  writes
• 5:36
  attacker-controlled
• 5:37
  bytes
• 5:37
  past
• 5:38
  the
• 5:38
  buffer.
• 5:39
  An
• 5:39
  unauthenticated
• 5:40
  attacker
• 5:40
  on
• 5:40
  the
• 5:40
  local
• 5:40
  network
• 5:41
  can
• 5:41
  trigger
• 5:41
  it
• 5:42
  with
• 5:42
  a
• 5:42
  unicast
• 5:42
  ICMP
• 5:43
  error
• 5:43
  during
• 5:44
  the
• 5:44
  boot
• 5:44
  window,
• 5:45
  yielding
• 5:45
  denial
• 5:46
  of
• 5:46
  provisioning
• 5:46
  or,
• 5:46
  against
• 5:47
  a
• 5:47
  known
• 5:47
  firmware
• 5:47
  build,
• 5:48
  code
• 5:48
  execution
• 5:49
  in
• 5:49
  ring
• 5:49
  0
• 5:50
  before
• 5:50
  Secure
• 5:50
  Boot
• 5:51
  or
• 5:51
  the
• 5:51
  TPM
• 5:51
  has
• 5:52
  measured
• 5:52
  anything
• 5:52
  .
• 5:53
  The
• 5:53
  discoverers
• 5:54
  are
• 5:54
  Alex
• 5:54
  Matrosov
• 5:55
  of
• 5:55
  Anthropic
• 5:56
  and
• 5:56
  Dan
• 5:56
  Holmes
• 5:56
  of
• 5:56
  Google,
• 5:57
  which
• 5:57
  quietly
• 5:58
  connects
• 5:58
  this
• 5:58
  to
• 5:58
  the
• 5:59
  first
• 5:59
  theme:
• 6:00
  the
• 6:00
  same
• 6:00
  vendor
• 6:00
  publishing
• 6:01
  exploit-chain
• 6:02
  research
• 6:02
  is
• 6:03
  also
• 6:03
  feeding
• 6:03
  firmware
• 6:04
  bugs
• 6:04
  into
• 6:04
  coordinated
• 6:05
  disclosure.
• 6:06
  That
• 6:07
  disclosure
• 6:07
  got
• 6:07
  messy
• 6:08
  when
• 6:08
  AMI
• 6:09
  published
• 6:09
  the
• 6:09
  bug
• 6:09
  with
• 6:09
  a
• 6:09
  code
• 6:10
  fix
• 6:10
  ahead
• 6:10
  of
• 6:10
  a
• 6:10
  planned
• 6:11
  January
• 6:11
  timeline,
• 6:12
  forcing
• 6:13
  TianoCore
• 6:13
  to
• 6:13
  issue
• 6:14
  its
• 6:14
  advisory
• 6:14
  early.
• 6:15
  HD
• 6:15
  Moore
• 6:16
  warned
• 6:16
  the
• 6:16
  flaw
• 6:16
  could
• 6:16
  hijack
• 6:17
  program
• 6:17
  flow
• 6:17
  while
• 6:18
  the
• 6:18
  boot
• 6:18
  process
• 6:18
  still
• 6:19
  has
• 6:19
  access
• 6:19
  to
• 6:19
  the
• 6:19
  TPM,
• 6:20
  which
• 6:20
  "could
• 6:20
  be
• 6:21
  used
• 6:21
  to
• 6:21
  break
• 6:21
  disk
• 6:21
  encryption,"
• 6:22
  and
• 6:22
  called
• 6:23
  it
• 6:23
  a
• 6:23
  likely
• 6:23
  "forever"
• 6:24
  bug
• 6:24
  in
• 6:24
  end-of-support
• 6:25
  hardware.
• 6:26
  The
• 6:26
  one
• 6:26
  piece
• 6:26
  of
• 6:27
  good
• 6:27
  news
• 6:27
  comes
• 6:27
  from
• 6:27
  Eclypsium's
• 6:28
  Jesse
• 6:28
  Michael:
• 6:29
  unlike
• 6:29
  BootHole,
• 6:30
  BlackLotus,
• 6:31
  or
• 6:31
  Baton
• 6:31
  Drop,
• 6:32
  this
• 6:32
  one
• 6:32
  needs
• 6:33
  only
• 6:33
  a
• 6:33
  BIOS
• 6:34
  update,
• 6:34
  with
• 6:34
  no
• 6:35
  key
• 6:35
  revocation
• 6:36
  or
• 6:36
  dbx
• 6:36
  management.
• 6:38
  It
• 6:38
  does
• 6:38
  not
• 6:38
  touch
• 6:38
  home
• 6:39
  users
• 6:39
  booting
• 6:39
  from
• 6:39
  disk;
• 6:40
  it
• 6:40
  threatens
• 6:41
  network-booting
• 6:41
  servers
• 6:42
  with
• 6:42
  PXE
• 6:43
  enabled,
• 6:43
  which
• 6:43
  is
• 6:44
  to
• 6:44
  say
• 6:44
  most
• 6:44
  of
• 6:44
  them
• 6:44
  .
• 6:45
  Further
• 6:46
  down
• 6:46
  the
• 6:46
  stack,
• 6:46
  three
• 6:47
  independent
• 6:47
  TLS
• 6:48
  bugs
• 6:48
  in
• 6:48
  GoAhead,
• 6:49
  the
• 6:49
  embedded
• 6:50
  web
• 6:50
  server
• 6:50
  in
• 6:51
  hundreds
• 6:51
  of
• 6:51
  millions
• 6:51
  of
• 6:52
  IoT,
• 6:52
  OT,
• 6:53
  and
• 6:53
  military
• 6:54
  devices
• 6:54
  from
• 6:54
  Palo
• 6:55
  Alto
• 6:55
  Networks
• 6:55
  to
• 6:55
  Boeing,
• 6:56
  show
• 6:56
  how
• 6:57
  configuration
• 6:57
  decides
• 6:58
  exposure.
• 6:59
  The
• 6:59
  OpenSSL-compatible
• 7:01
  backend
• 7:01
  loads
• 7:01
  a
• 7:01
  certificate
• 7:02
  revocation
• 7:03
  list
• 7:03
  but
• 7:03
  never
• 7:04
  enables
• 7:04
  revocation
• 7:05
  checking,
• 7:05
  so
• 7:06
  revoked
• 7:06
  client
• 7:06
  certificates
• 7:07
  sail
• 7:07
  through
• 7:07
  mutual
• 7:08
  TLS
• 7:08
  at
• 7:09
  CVSS
• 7:09
  9.1;
• 7:11
  a
• 7:11
  LibreSSL
• 7:11
  legacy
• 7:12
  verifier
• 7:13
  accepts
• 7:13
  over-depth
• 7:13
  certificate
• 7:14
  chains;
• 7:15
  and
• 7:15
  an
• 7:15
  Mbed
• 7:15
  TLS
• 7:16
  one-byte
• 7:16
  typo
• 7:17
  truncates
• 7:17
  a
• 7:17
  2048-bit
• 7:19
  Diffie-Hellman
• 7:19
  modulus
• 7:20
  to
• 7:20
  a
• 7:20
  single
• 7:20
  byte,
• 7:21
  leaving
• 7:21
  only
• 7:22
  eight
• 7:22
  possible
• 7:22
  shared
• 7:23
  secrets
• 7:23
  and
• 7:23
  making
• 7:23
  TLS
• 7:24
  1.2
• 7:25
  DHE
• 7:25
  RSA
• 7:26
  sessions
• 7:26
  decryptable
• 7:27
  by
• 7:27
  a
• 7:27
  passive
• 7:27
  observer
• 7:27
  .
• 7:28
  And
• 7:28
  the
• 7:29
  PHP
• 7:29
  megabug
• 7:30
  has
• 7:30
  stopped
• 7:30
  being
• 7:30
  theoretical.
• 7:32
  A
• 7:32
  honeypot
• 7:32
  report
• 7:33
  sealed
• 7:33
  35
• 7:34
  confirmed
• 7:34
  remote
• 7:34
  code
• 7:35
  executions
• 7:35
  on
• 7:36
  October
• 7:36
  24
• 7:37
  across
• 7:37
  disposable
• 7:37
  PHP
• 7:38
  sensors,
• 7:39
  with
• 7:39
  a
• 7:39
  shell
• 7:39
  parented
• 7:40
  by
• 7:40
  Apache
• 7:40
  workers
• 7:41
  as
• 7:41
  the
• 7:41
  only
• 7:41
  success
• 7:42
  criterion,
• 7:43
  driven
• 7:43
  by
• 7:43
  a
• 7:43
  four-stage
• 7:44
  multipart
• 7:44
  POST
• 7:44
  chain
• 7:45
  that
• 7:45
  leaks
• 7:45
  memory
• 7:46
  to
• 7:46
  defeat
• 7:46
  address
• 7:47
  randomization
• 7:47
  before
• 7:48
  adjusting
• 7:48
  its
• 7:48
  payload
• 7:49
  per
• 7:49
  target.
• 7:49
  The
• 7:50
  authors
• 7:50
  are
• 7:50
  scrupulous
• 7:51
  about
• 7:51
  limits,
• 7:52
  calling
• 7:52
  it
• 7:52
  "two
• 7:52
  source-address
• 7:53
  episodes
• 7:53
  using
• 7:54
  closely
• 7:54
  related
• 7:55
  tooling"
• 7:55
  rather
• 7:55
  than
• 7:55
  attribution,
• 7:56
  and
• 7:57
  noting
• 7:57
  the
• 7:57
  single-worker
• 7:58
  setup
• 7:58
  was
• 7:58
  a
• 7:58
  deliberately
• 7:59
  favorable
• 7:59
  profile,
• 8:00
  not
• 8:00
  a
• 8:00
  reliability
• 8:01
  claim
• 8:01
  .
• 8:07
  Anthropic,
• 8:08
  the
• 8:08
  company
• 8:08
  that
• 8:08
  built
• 8:09
  its
• 8:09
  brand
• 8:09
  on
• 8:09
  safety,
• 8:10
  shipped
• 8:10
  a
• 8:10
  way
• 8:11
  to
• 8:11
  throw
• 8:11
  away
• 8:11
  Claude
• 8:11
  Code's
• 8:12
  last
• 8:12
  guardrail.
• 8:13
  The
• 8:13
  Register
• 8:14
  reports
• 8:14
  that
• 8:14
  the
• 8:14
  latest
• 8:15
  build
• 8:15
  adds
• 8:15
  an
• 8:16
  environment
• 8:16
  variable
• 8:17
  that
• 8:17
  lets
• 8:17
  the
• 8:17
  agent
• 8:17
  run
• 8:18
  a
• 8:18
  file-deleting
• 8:18
  command
• 8:19
  without
• 8:19
  asking,
• 8:20
  closing
• 8:20
  the
• 8:20
  gap
• 8:21
  the
• 8:21
  existing
• 8:21
  bypass
• 8:22
  mode
• 8:22
  had
• 8:22
  left
• 8:22
  open.
• 8:23
  Daisy
• 8:24
  Hollman
• 8:24
  of
• 8:24
  Anthropic
• 8:25
  announced
• 8:25
  it:
• 8:26
  the
• 8:26
  existing
• 8:26
  flag
• 8:27
  "doesn't
• 8:27
  actually
• 8:28
  approve
• 8:28
  EVERYTHING,
• 8:29
  we
• 8:29
  still
• 8:29
  block
• 8:30
  some
• 8:30
  obvious
• 8:30
  things,"
• 8:31
  she
• 8:31
  wrote,
• 8:32
  "but
• 8:32
  today's
• 8:32
  release
• 8:33
  will
• 8:33
  let
• 8:33
  you
• 8:33
  disable
• 8:34
  this
• 8:34
  if
• 8:34
  you
• 8:34
  really
• 8:34
  want."
• 8:35
  The
• 8:35
  change
• 8:36
  is
• 8:36
  absent
• 8:36
  from
• 8:36
  the
• 8:36
  v2.1.311
• 8:40
  changelog,
• 8:41
  and
• 8:41
  the
• 8:41
  article
• 8:41
  captures
• 8:42
  the
• 8:42
  house
• 8:42
  skepticism
• 8:43
  well,
• 8:43
  noting
• 8:44
  the
• 8:44
  documented
• 8:44
  cases
• 8:45
  of
• 8:45
  agents
• 8:45
  issuing
• 8:46
  exactly
• 8:46
  this
• 8:46
  kind
• 8:47
  of
• 8:47
  command
• 8:47
  and
• 8:47
  destroying
• 8:48
  data
• 8:48
  .
• 8:49
  The
• 8:49
  right
• 8:49
  way
• 8:49
  to
• 8:49
  read
• 8:50
  this
• 8:50
  is
• 8:50
  not
• 8:50
  as
• 8:51
  a
• 8:51
  vulnerability
• 8:51
  but
• 8:51
  as
• 8:52
  a
• 8:52
  configuration
• 8:52
  option,
• 8:53
  and
• 8:53
  the
• 8:53
  recurring
• 8:54
  lesson
• 8:54
  applies:
• 8:55
  if
• 8:55
  a
• 8:55
  user
• 8:55
  can
• 8:55
  toggle
• 8:56
  the
• 8:56
  restraint
• 8:56
  off
• 8:57
  from
• 8:57
  inside
• 8:57
  the
• 8:57
  agent's
• 8:58
  own
• 8:58
  config,
• 8:59
  the
• 8:59
  restraint
• 8:59
  never
• 9:00
  belonged
• 9:00
  there.
• 9:01
  Enforcement
• 9:02
  has
• 9:02
  to
• 9:02
  live
• 9:02
  outside,
• 9:03
  in
• 9:03
  hooks,
• 9:03
  managed
• 9:04
  settings,
• 9:05
  or
• 9:05
  a
• 9:05
  kernel
• 9:05
  sandbox.
• 9:07
  The
• 9:07
  commercial
• 9:07
  market
• 9:07
  has
• 9:07
  already
• 9:08
  productized
• 9:08
  that
• 9:09
  logic
• 9:09
  in
• 9:09
  both
• 9:09
  directions.
• 9:11
  Refuseless
• 9:11
  sells
• 9:11
  access
• 9:12
  to
• 9:12
  abliterated
• 9:13
  open
• 9:13
  models,
• 9:13
  GLM-5.3
• 9:15
  and
• 9:15
  DeepSeek
• 9:16
  V4.1
• 9:18
  Flash
• 9:18
  among
• 9:18
  them,
• 9:19
  served
• 9:19
  through
• 9:19
  OpenAI-
• 9:20
  and
• 9:20
  Anthropic-compatible
• 9:21
  APIs
• 9:22
  with
• 9:22
  zero
• 9:22
  retention
• 9:23
  and
• 9:23
  pitched
• 9:23
  directly
• 9:23
  at
• 9:24
  security
• 9:24
  teams,
• 9:25
  complete
• 9:25
  with
• 9:25
  CyberGym
• 9:26
  and
• 9:26
  ExploitGym
• 9:26
  scores
• 9:27
  on
• 9:27
  the
• 9:27
  model
• 9:27
  cards.
• 9:28
  Its
• 9:29
  own
• 9:29
  copy
• 9:29
  is
• 9:29
  refreshingly
• 9:30
  honest
• 9:30
  about
• 9:31
  what
• 9:31
  it
• 9:31
  is
• 9:31
  buying
• 9:31
  you:
• 9:32
  abliteration
• 9:33
  edits
• 9:33
  the
• 9:33
  refusal
• 9:34
  direction
• 9:34
  out
• 9:34
  of
• 9:34
  the
• 9:34
  weights,
• 9:35
  so
• 9:35
  it
• 9:35
  "removes
• 9:36
  a
• 9:36
  brake,"
• 9:36
  adds
• 9:37
  no
• 9:37
  knowledge,
• 9:38
  and
• 9:38
  raises
• 9:38
  no
• 9:38
  benchmark
• 9:39
  .
• 9:40
  The
• 9:40
  same
• 9:40
  current
• 9:40
  surfaces
• 9:41
  on
• 9:41
  Hugging
• 9:41
  Face,
• 9:42
  where
• 9:42
  an
• 9:42
  FP8
• 9:43
  uncensored
• 9:43
  build
• 9:44
  of
• 9:44
  Qwen3.8-27B
• 9:48
  trended
• 9:48
  this
• 9:49
  week
• 9:49
  .
• 9:50
  On
• 9:50
  the
• 9:50
  defensive
• 9:50
  flank,
• 9:51
  aitotal.io
• 9:52
  from
• 9:52
  Zenity
• 9:53
  is
• 9:53
  building
• 9:53
  a
• 9:53
  VirusTotal
• 9:54
  for
• 9:54
  agent
• 9:54
  skills:
• 9:55
  it
• 9:55
  runs
• 9:56
  a
• 9:56
  SKILL.md
• 9:57
  or
• 9:57
  package
• 9:57
  in
• 9:58
  an
• 9:58
  isolated
• 9:58
  sandbox,
• 9:59
  records
• 10:00
  every
• 10:00
  network,
• 10:01
  filesystem,
• 10:01
  and
• 10:01
  process
• 10:02
  call
• 10:02
  across
• 10:03
  install,
• 10:03
  load,
• 10:04
  and
• 10:04
  use,
• 10:04
  and
• 10:05
  returns
• 10:05
  a
• 10:05
  benign,
• 10:06
  suspicious,
• 10:07
  or
• 10:07
  malicious
• 10:07
  verdict
• 10:08
  backed
• 10:08
  by
• 10:08
  behavior.
• 10:09
  Its
• 10:09
  sample
• 10:10
  report
• 10:10
  caught
• 10:10
  an
• 10:11
  "audit"
• 10:11
  skill
• 10:11
  quietly
• 10:12
  uploading
• 10:13
  a
• 10:13
  bundled
• 10:13
  file
• 10:13
  to
• 10:14
  its
• 10:14
  author's
• 10:14
  own
• 10:14
  endpoint,
• 10:15
  the
• 10:15
  kind
• 10:16
  of
• 10:16
  rug-pull
• 10:16
  that
• 10:16
  static
• 10:17
  review
• 10:17
  cannot
• 10:17
  see
• 10:18
  .
• 10:18
  Strip
• 10:19
  the
• 10:19
  refusals
• 10:19
  on
• 10:19
  one
• 10:20
  side,
• 10:20
  detonate
• 10:21
  the
• 10:21
  payloads
• 10:21
  on
• 10:21
  the
• 10:22
  other.
• 10:26
  The
• 10:26
  most
• 10:27
  consequential
• 10:27
  paper
• 10:28
  of
• 10:28
  the
• 10:28
  week
• 10:28
  is
• 10:28
  also
• 10:29
  the
• 10:29
  most
• 10:29
  carefully
• 10:30
  hedged,
• 10:30
  and
• 10:31
  the
• 10:31
  gap
• 10:31
  between
• 10:31
  the
• 10:31
  document
• 10:32
  and
• 10:32
  its
• 10:32
  viral
• 10:33
  retelling
• 10:33
  is
• 10:33
  itself
• 10:34
  the
• 10:34
  story.
• 10:35
  The
• 10:35
  preprint,
• 10:36
  titled
• 10:36
  "The
• 10:37
  Exponential
• 10:37
  Time
• 10:37
  Hypothesis
• 10:38
  Is
• 10:38
  False,"
• 10:39
  is
• 10:39
  hosted
• 10:40
  under
• 10:40
  OpenAI's
• 10:40
  math
• 10:41
  repository,
• 10:42
  dated
• 10:42
  October
• 10:43
  17
• 10:43
  and
• 10:44
  revised
• 10:44
  October
• 10:45
  25,
• 10:46
  and
• 10:46
  lists
• 10:46
  its
• 10:46
  author
• 10:47
  as
• 10:47
  nothing
• 10:47
  more
• 10:47
  specific
• 10:48
  than
• 10:48
  "OPENAI"
• 10:51
  across
• 10:51
  67
• 10:52
  pages.
• 10:53
  Theorem
• 10:54
  1.1
• 10:55
  claims
• 10:55
  a
• 10:55
  deterministic
• 10:56
  algorithm
• 10:56
  that
• 10:57
  counts
• 10:57
  satisfying
• 10:58
  assignments
• 10:58
  of
• 10:58
  a
• 10:58
  k-CNF
• 10:59
  formula
• 11:00
  in
• 11:00
  time
• 11:00
  scaling
• 11:00
  as
• 11:01
  two
• 11:01
  to
• 11:01
  the
• 11:01
  power
• 11:01
  of
• 11:02
  roughly
• 11:02
  n
• 11:02
  log
• 11:03
  log
• 11:03
  n
• 11:03
  over
• 11:03
  the
• 11:03
  square
• 11:04
  root
• 11:04
  of
• 11:04
  log
• 11:04
  n,
• 11:05
  which
• 11:05
  is
• 11:05
  subexponential,
• 11:07
  and
• 11:07
  therefore
• 11:07
  refutes
• 11:08
  the
• 11:08
  Exponential
• 11:08
  Time
• 11:09
  Hypothesis
• 11:10
  along
• 11:10
  with
• 11:10
  its
• 11:10
  strong,
• 11:11
  gap,
• 11:11
  and
• 11:12
  counting
• 11:12
  variants.
• 11:13
  The
• 11:13
  proof
• 11:14
  writes
• 11:14
  the
• 11:14
  solution
• 11:14
  count
• 11:15
  as
• 11:15
  a
• 11:15
  sum
• 11:15
  over
• 11:16
  all
• 11:16
  assignments
• 11:17
  of
• 11:17
  a
• 11:17
  product
• 11:17
  of
• 11:17
  clause
• 11:17
  polynomials,
• 11:19
  introduces
• 11:19
  replacement
• 11:20
  linear
• 11:20
  forms
• 11:21
  that
• 11:21
  recover
• 11:21
  the
• 11:21
  original
• 11:22
  variables
• 11:22
  at
• 11:22
  a
• 11:22
  limit
• 11:23
  point,
• 11:23
  and
• 11:23
  uses
• 11:24
  an
• 11:24
  algebraic
• 11:24
  design
• 11:25
  that
• 11:25
  lets
• 11:25
  the
• 11:25
  sum
• 11:26
  factor
• 11:26
  across
• 11:26
  small
• 11:27
  groups
• 11:27
  of
• 11:27
  variables
• 11:28
  before
• 11:28
  recovering
• 11:28
  the
• 11:29
  answer
• 11:29
  by
• 11:29
  Hermite
• 11:29
  interpolation
• 11:30
  .
• 11:31
  What
• 11:31
  settles
• 11:31
  whether
• 11:32
  this
• 11:32
  matters
• 11:32
  is
• 11:32
  verification,
• 11:33
  and
• 11:34
  here
• 11:34
  the
• 11:34
  primary
• 11:34
  document
• 11:35
  and
• 11:35
  the
• 11:35
  public
• 11:35
  narrative
• 11:36
  diverge
• 11:36
  sharply.
• 11:37
  The
• 11:38
  widely
• 11:38
  circulated
• 11:39
  explanation
• 11:39
  says
• 11:40
  an
• 11:40
  internal
• 11:40
  OpenAI
• 11:41
  model
• 11:41
  found
• 11:41
  the
• 11:41
  algorithm
• 11:42
  and
• 11:42
  checked
• 11:43
  the
• 11:43
  proof
• 11:43
  in
• 11:43
  Lean,
• 11:44
  and
• 11:44
  that
• 11:44
  Russell
• 11:45
  Impagliazzo,
• 11:46
  Ramamohan
• 11:46
  Paturi,
• 11:47
  and
• 11:48
  MIT's
• 11:48
  Ryan
• 11:48
  Williams,
• 11:49
  including
• 11:50
  two
• 11:50
  of
• 11:50
  the
• 11:50
  hypothesis's
• 11:51
  original
• 11:52
  authors,
• 11:52
  have
• 11:52
  been
• 11:53
  asked
• 11:53
  to
• 11:53
  organize
• 11:54
  an
• 11:54
  independent
• 11:54
  examination
• 11:55
  .
• 11:56
  The
• 11:56
  PDF
• 11:56
  itself
• 11:57
  says
• 11:57
  none
• 11:57
  of
• 11:57
  that.
• 11:58
  It
• 11:58
  contains
• 11:58
  no
• 11:58
  mention
• 11:59
  of
• 11:59
  a
• 11:59
  language
• 11:59
  model,
• 12:00
  no
• 12:00
  mention
• 12:00
  of
• 12:01
  Lean
• 12:01
  or
• 12:01
  any
• 12:01
  formal
• 12:02
  verification,
• 12:03
  and
• 12:03
  no
• 12:03
  human
• 12:03
  names,
• 12:04
  which
• 12:04
  is
• 12:05
  a
• 12:05
  conspicuous
• 12:05
  departure
• 12:06
  from
• 12:06
  every
• 12:06
  prior
• 12:07
  OpenAI
• 12:07
  mathematics
• 12:08
  claim,
• 12:08
  each
• 12:09
  of
• 12:09
  which
• 12:09
  shipped
• 12:09
  a
• 12:09
  Lean
• 12:10
  artifact
• 12:10
  alongside
• 12:11
  it.
• 12:12
  The
• 12:12
  paper
• 12:12
  also
• 12:12
  states
• 12:13
  its
• 12:13
  own
• 12:13
  limits
• 12:13
  plainly:
• 12:14
  the
• 12:14
  result
• 12:15
  does
• 12:15
  not
• 12:15
  transfer
• 12:16
  to
• 12:16
  all
• 12:16
  of
• 12:16
  NP
• 12:17
  under
• 12:17
  general
• 12:17
  reductions,
• 12:18
  and
• 12:18
  the
• 12:18
  algorithm
• 12:19
  is
• 12:19
  still
• 12:19
  slow
• 12:20
  in
• 12:20
  absolute
• 12:20
  terms.
• 12:21
  A
• 12:21
  sharp
• 12:22
  reply
• 12:22
  to
• 12:22
  the
• 12:22
  viral
• 12:22
  thread
• 12:23
  made
• 12:23
  the
• 12:23
  operative
• 12:24
  point,
• 12:24
  that
• 12:24
  a
• 12:24
  non-constructive
• 12:25
  argument
• 12:26
  can
• 12:26
  show
• 12:26
  such
• 12:26
  an
• 12:26
  algorithm
• 12:27
  exists
• 12:27
  in
• 12:28
  principle
• 12:28
  without
• 12:28
  anyone
• 12:29
  having
• 12:29
  one
• 12:29
  to
• 12:29
  run.
• 12:30
  If
• 12:31
  the
• 12:31
  refutation
• 12:31
  holds,
• 12:32
  the
• 12:32
  large
• 12:33
  body
• 12:33
  of
• 12:33
  fine-grained
• 12:34
  complexity
• 12:34
  results
• 12:35
  that
• 12:35
  rest
• 12:35
  on
• 12:35
  ETH
• 12:36
  as
• 12:36
  a
• 12:36
  measuring
• 12:36
  stick
• 12:37
  loses
• 12:37
  its
• 12:37
  foundation.
• 12:39
  Until
• 12:39
  an
• 12:39
  independent
• 12:40
  panel
• 12:40
  reports
• 12:40
  and
• 12:41
  a
• 12:41
  formalization
• 12:41
  appears,
• 12:42
  the
• 12:42
  honest
• 12:43
  status
• 12:43
  is
• 12:43
  a
• 12:43
  serious,
• 12:44
  checkable
• 12:44
  claim
• 12:45
  awaiting
• 12:45
  the
• 12:45
  scrutiny
• 12:46
  the
• 12:46
  founders
• 12:46
  were
• 12:46
  asked
• 12:47
  to
• 12:47
  lead.
• 12:51
  Steve
• 12:52
  Yegge's
• 12:52
  "The
• 12:52
  Counting
• 12:53
  House"
• 12:53
  is
• 12:53
  the
• 12:53
  counterweight
• 12:54
  to
• 12:54
  his
• 12:54
  own
• 12:54
  agent-swarm
• 12:55
  software
• 12:56
  factory,
• 12:57
  and
• 12:57
  it
• 12:57
  is
• 12:57
  the
• 12:57
  more
• 12:57
  disciplined
• 12:58
  design.
• 12:59
  Where
• 12:59
  his
• 12:59
  Wheelhouse
• 12:59
  treats
• 13:00
  fifty
• 13:00
  agents
• 13:01
  as
• 13:01
  long-lived
• 13:01
  colleagues
• 13:02
  producing
• 13:02
  thousands
• 13:03
  of
• 13:03
  commits
• 13:03
  a
• 13:03
  week,
• 13:04
  the
• 13:04
  Counting
• 13:04
  House
• 13:05
  does
• 13:05
  the
• 13:05
  opposite
• 13:06
  for
• 13:06
  high-volume
• 13:06
  human
• 13:07
  casework:
• 13:07
  stateless
• 13:08
  one-shot
• 13:09
  models,
• 13:09
  mechanical
• 13:10
  and
• 13:10
  deterministic,
• 13:11
  with
• 13:11
  only
• 13:11
  the
• 13:12
  scopes
• 13:12
  they
• 13:12
  need,
• 13:13
  no
• 13:13
  credentials
• 13:13
  of
• 13:14
  their
• 13:14
  own,
• 13:14
  no
• 13:15
  memory,
• 13:15
  and
• 13:15
  never
• 13:16
  permission
• 13:16
  to
• 13:16
  write
• 13:16
  to
• 13:17
  a
• 13:17
  system
• 13:17
  of
• 13:17
  record.
• 13:18
  The
• 13:18
  work
• 13:18
  is
• 13:18
  tractable
• 13:19
  because
• 13:19
  a
• 13:19
  company
• 13:20
  handling
• 13:20
  tens
• 13:21
  of
• 13:21
  millions
• 13:21
  of
• 13:21
  contacts
• 13:22
  a
• 13:22
  year
• 13:22
  still
• 13:22
  has
• 13:22
  only
• 13:23
  a
• 13:23
  few
• 13:23
  dozen
• 13:23
  to
• 13:23
  a
• 13:24
  few
• 13:24
  hundred
• 13:24
  kinds
• 13:24
  of
• 13:24
  case,
• 13:25
  which
• 13:25
  he
• 13:25
  calls
• 13:25
  intents.
• 13:27
  The
• 13:27
  core
• 13:27
  is
• 13:27
  a
• 13:27
  policy
• 13:28
  table,
• 13:28
  one
• 13:29
  row
• 13:29
  per
• 13:29
  intent,
• 13:30
  specifying
• 13:30
  which
• 13:31
  model
• 13:31
  may
• 13:31
  handle
• 13:31
  it
• 13:31
  at
• 13:32
  which
• 13:32
  of
• 13:32
  five
• 13:32
  autonomy
• 13:33
  tiers,
• 13:33
  from
• 13:33
  off
• 13:34
  through
• 13:34
  shadow,
• 13:34
  assist,
• 13:35
  and
• 13:35
  contain
• 13:36
  up
• 13:36
  to
• 13:36
  transact,
• 13:37
  with
• 13:37
  what
• 13:37
  tools
• 13:37
  and
• 13:38
  which
• 13:38
  human
• 13:38
  approvers.
• 13:39
  Models
• 13:39
  earn
• 13:39
  promotion
• 13:40
  through
• 13:40
  evals
• 13:40
  built
• 13:41
  on
• 13:41
  gold
• 13:41
  sets
• 13:41
  of
• 13:42
  200
• 13:42
  to
• 13:42
  1000
• 13:43
  human-handled
• 13:44
  cases,
• 13:44
  scored
• 13:44
  blind,
• 13:45
  where
• 13:45
  an
• 13:46
  LLM
• 13:46
  judge
• 13:46
  that
• 13:46
  was
• 13:47
  never
• 13:47
  tuned
• 13:47
  on
• 13:47
  the
• 13:47
  set
• 13:48
  must
• 13:48
  agree
• 13:48
  with
• 13:48
  humans
• 13:49
  about
• 13:49
  as
• 13:49
  often
• 13:49
  as
• 13:49
  humans
• 13:50
  agree
• 13:50
  with
• 13:50
  each
• 13:50
  other,
• 13:51
  typically
• 13:51
  around
• 13:52
  90%
• 13:52
  .
• 13:53
  His
• 13:53
  sharpest
• 13:54
  operational
• 13:54
  point
• 13:55
  is
• 13:55
  to
• 13:55
  watch
• 13:55
  the
• 13:55
  confusion
• 13:56
  matrix
• 13:56
  rather
• 13:56
  than
• 13:57
  raw
• 13:57
  accuracy,
• 13:58
  because
• 13:58
  mistake
• 13:58
  costs
• 13:59
  are
• 13:59
  wildly
• 13:59
  asymmetric:
• 14:00
  a
• 14:01
  misroute
• 14:01
  costs
• 14:01
  a
• 14:01
  few
• 14:02
  dollars,
• 14:03
  while
• 14:03
  stranding
• 14:03
  a
• 14:03
  child
• 14:04
  at
• 14:04
  an
• 14:04
  airport
• 14:04
  costs
• 14:05
  $25,000
• 14:06
  and
• 14:07
  a
• 14:07
  letter
• 14:07
  to
• 14:07
  a
• 14:07
  regulator.
• 14:08
  Money
• 14:09
  intents
• 14:09
  want
• 14:09
  precision;
• 14:10
  special-assistance
• 14:11
  intents
• 14:11
  want
• 14:12
  near-total
• 14:12
  recall.
• 14:13
  The
• 14:13
  orchestrator
• 14:14
  that
• 14:14
  validates
• 14:15
  write
• 14:15
  intents
• 14:15
  against
• 14:15
  live
• 14:16
  systems
• 14:16
  is
• 14:17
  mechanical,
• 14:17
  built
• 14:18
  on
• 14:18
  the
• 14:18
  Saga
• 14:18
  pattern,
• 14:19
  and
• 14:19
  his
• 14:19
  closing
• 14:20
  warning
• 14:20
  is
• 14:20
  the
• 14:20
  one
• 14:21
  to
• 14:21
  tattoo
• 14:21
  on
• 14:22
  a
• 14:22
  wall:
• 14:22
  whoever
• 14:23
  owns
• 14:23
  the
• 14:23
  evals
• 14:24
  decides
• 14:24
  which
• 14:24
  models
• 14:25
  you
• 14:25
  can
• 14:25
  switch
• 14:25
  to,
• 14:26
  so
• 14:26
  make
• 14:26
  sure
• 14:26
  it
• 14:26
  is
• 14:27
  your
• 14:27
  own
• 14:27
  data-science
• 14:28
  team.
• 14:29
  That
• 14:29
  same
• 14:29
  precision-over-recall
• 14:30
  philosophy
• 14:31
  shows
• 14:31
  up
• 14:31
  in
• 14:31
  Alibaba's
• 14:32
  open-code-review,
• 14:33
  the
• 14:33
  company's
• 14:34
  internal
• 14:34
  AI
• 14:35
  review
• 14:35
  assistant
• 14:35
  now
• 14:36
  released
• 14:36
  as
• 14:36
  a
• 14:36
  Go
• 14:36
  CLI.
• 14:37
  Its
• 14:37
  pitch
• 14:38
  is
• 14:38
  a
• 14:38
  direct
• 14:38
  shot
• 14:39
  at
• 14:39
  pure-LLM
• 14:40
  reviewers
• 14:40
  like
• 14:40
  Claude
• 14:40
  Code
• 14:41
  with
• 14:41
  Skills,
• 14:41
  which
• 14:42
  it
• 14:42
  argues
• 14:42
  suffer
• 14:43
  incomplete
• 14:43
  coverage,
• 14:44
  position
• 14:44
  drift,
• 14:44
  and
• 14:45
  unstable
• 14:45
  quality
• 14:46
  because
• 14:46
  a
• 14:46
  purely
• 14:47
  language-driven
• 14:47
  process
• 14:48
  lacks
• 14:48
  hard
• 14:48
  constraints.
• 14:49
  Open
• 14:50
  Code
• 14:50
  Review
• 14:50
  instead
• 14:51
  hard-codes
• 14:51
  the
• 14:51
  steps
• 14:52
  that
• 14:52
  must
• 14:52
  not
• 14:52
  go
• 14:53
  wrong,
• 14:53
  file
• 14:53
  selection,
• 14:54
  bundling,
• 14:55
  and
• 14:55
  rule
• 14:55
  matching,
• 14:56
  and
• 14:56
  lets
• 14:56
  the
• 14:56
  agent
• 14:57
  handle
• 14:57
  only
• 14:57
  the
• 14:57
  dynamic
• 14:58
  judgment.
• 14:59
  On
• 14:59
  its
• 14:59
  own
• 14:59
  AACR-Bench
• 15:01
  of
• 15:01
  200
• 15:02
  real
• 15:02
  pull
• 15:02
  requests,
• 15:03
  it
• 15:03
  claims
• 15:04
  higher
• 15:04
  precision
• 15:04
  and
• 15:04
  F1
• 15:05
  than
• 15:05
  general-purpose
• 15:05
  agents
• 15:06
  on
• 15:06
  the
• 15:06
  same
• 15:06
  underlying
• 15:07
  model
• 15:07
  while
• 15:07
  burning
• 15:07
  roughly
• 15:08
  a
• 15:08
  ninth
• 15:08
  of
• 15:08
  the
• 15:08
  tokens,
• 15:09
  with
• 15:09
  deliberately
• 15:10
  lower
• 15:10
  recall
• 15:10
  as
• 15:11
  the
• 15:11
  trade
• 15:11
  .
• 15:11
  NVIDIA's
• 15:12
  Kumo
• 15:12
  Relational
• 15:13
  pushes
• 15:13
  the
• 15:13
  same
• 15:13
  enterprise-systems
• 15:14
  theme
• 15:14
  into
• 15:15
  prediction:
• 15:16
  a
• 15:16
  pretrained
• 15:16
  relational
• 15:17
  foundation
• 15:17
  model
• 15:18
  that
• 15:18
  runs
• 15:18
  directly
• 15:19
  on
• 15:19
  a
• 15:19
  database,
• 15:20
  gathering
• 15:21
  historical
• 15:21
  rows
• 15:22
  as
• 15:22
  in-context
• 15:22
  examples
• 15:23
  to
• 15:23
  predict
• 15:23
  churn
• 15:24
  or
• 15:24
  fraud
• 15:24
  in
• 15:24
  a
• 15:24
  single
• 15:25
  forward
• 15:25
  pass
• 15:26
  with
• 15:26
  no
• 15:26
  task-specific
• 15:27
  training
• 15:28
  or
• 15:28
  feature
• 15:28
  engineering.
• 15:29
  Trained
• 15:30
  entirely
• 15:30
  on
• 15:30
  synthetic
• 15:31
  relational
• 15:31
  data,
• 15:32
  it
• 15:32
  beats
• 15:32
  gradient-boosted
• 15:33
  trees
• 15:34
  on
• 15:34
  expert
• 15:34
  features
• 15:35
  by
• 15:35
  nearly
• 15:35
  four
• 15:36
  AUROC
• 15:36
  points
• 15:36
  and
• 15:37
  XGBoost
• 15:37
  on
• 15:38
  SAP's
• 15:38
  benchmark
• 15:39
  by
• 15:39
  seven
• 15:39
  MRR
• 15:40
  points
• 15:40
  out
• 15:40
  of
• 15:40
  the
• 15:40
  box,
• 15:41
  while
• 15:41
  running
• 15:42
  about
• 15:42
  430
• 15:43
  times
• 15:43
  faster
• 15:43
  than
• 15:44
  a
• 15:44
  comparable
• 15:44
  baseline,
• 15:45
  though
• 15:45
  NVIDIA
• 15:46
  concedes
• 15:46
  that
• 15:47
  on
• 15:47
  very
• 15:47
  large
• 15:47
  training
• 15:48
  sets
• 15:48
  a
• 15:48
  dedicated
• 15:49
  supervised
• 15:49
  model
• 15:50
  may
• 15:50
  still
• 15:50
  edge
• 15:51
  it
• 15:51
  out
• 15:51
  .
• 15:56
  Liquid
• 15:56
  AI
• 15:56
  shipped
• 15:57
  not
• 15:57
  one
• 15:57
  small
• 15:57
  model
• 15:58
  but
• 15:58
  six,
• 15:58
  built
• 15:59
  on
• 15:59
  a
• 15:59
  new
• 15:59
  tri-hybrid
• 16:00
  architecture
• 16:00
  that
• 16:00
  mixes
• 16:01
  gated
• 16:01
  short
• 16:01
  convolutions
• 16:02
  with
• 16:02
  sliding-window
• 16:03
  and
• 16:03
  global
• 16:04
  attention.
• 16:05
  The
• 16:05
  anchor,
• 16:05
  LFM3-1.7B,
• 16:09
  carries
• 16:09
  a
• 16:09
  256K-token
• 16:11
  context,
• 16:12
  fits
• 16:12
  in
• 16:12
  about
• 16:12
  a
• 16:13
  gigabyte
• 16:13
  at
• 16:13
  4-bit,
• 16:14
  and
• 16:14
  is
• 16:14
  post-trained
• 16:15
  for
• 16:15
  agentic
• 16:15
  harnesses
• 16:16
  and
• 16:16
  tool
• 16:16
  calling,
• 16:17
  with
• 16:17
  the
• 16:17
  company
• 16:17
  candid
• 16:18
  that
• 16:18
  it
• 16:18
  is
• 16:18
  not
• 16:18
  meant
• 16:18
  for
• 16:19
  agentic
• 16:19
  coding
• 16:19
  or
• 16:20
  complex
• 16:20
  multi-step
• 16:21
  reasoning.
• 16:22
  The
• 16:22
  same
• 16:22
  release
• 16:22
  adds
• 16:23
  a
• 16:23
  2B
• 16:23
  vision-language
• 16:24
  model
• 16:24
  with
• 16:24
  video
• 16:25
  input,
• 16:25
  a
• 16:25
  2B
• 16:26
  audio
• 16:26
  model,
• 16:27
  a
• 16:27
  2B
• 16:27
  omni-modal
• 16:28
  model
• 16:28
  that
• 16:28
  replaces
• 16:29
  a
• 16:29
  separate
• 16:29
  transcription,
• 16:30
  vision,
• 16:31
  language,
• 16:31
  and
• 16:32
  speech
• 16:32
  stack,
• 16:33
  and
• 16:33
  two
• 16:33
  decision
• 16:33
  models
• 16:34
  that
• 16:34
  return
• 16:34
  calibrated
• 16:35
  answers
• 16:35
  in
• 16:36
  a
• 16:36
  single
• 16:36
  forward
• 16:36
  pass
• 16:37
  in
• 16:37
  under
• 16:37
  six
• 16:38
  milliseconds
• 16:38
  on
• 16:38
  an
• 16:39
  H100
• 16:39
  .
• 16:40
  On
• 16:40
  the
• 16:41
  document
• 16:41
  front,
• 16:41
  SoMark's
• 16:42
  OXR
• 16:43
  is
• 16:43
  a
• 16:43
  1.3-billion-parameter
• 16:45
  parser
• 16:45
  that
• 16:45
  claims
• 16:45
  the
• 16:46
  top
• 16:46
  spot
• 16:46
  on
• 16:46
  olmOCR-Bench,
• 16:47
  OmniDocBench,
• 16:48
  and
• 16:49
  DocParsingBench
• 16:50
  using
• 16:50
  a
• 16:50
  recursive
• 16:50
  self-improvement
• 16:51
  loop
• 16:51
  in
• 16:51
  post-training,
• 16:52
  with
• 16:53
  a
• 16:53
  pluggable
• 16:53
  layout
• 16:53
  module
• 16:54
  and
• 16:54
  recognition
• 16:55
  of
• 16:55
  thirteen
• 16:55
  element
• 16:56
  types
• 16:56
  down
• 16:56
  to
• 16:56
  chemical
• 16:57
  structures
• 16:57
  in
• 16:57
  SMILES,
• 16:58
  released
• 16:58
  under
• 16:58
  a
• 16:59
  custom
• 16:59
  Apache-derived
• 17:00
  license
• 17:00
  with
• 17:00
  a
• 17:00
  free
• 17:01
  API
• 17:01
  .
• 17:02
  The
• 17:02
  hardware
• 17:02
  subplot
• 17:03
  is
• 17:03
  noisier
• 17:04
  and
• 17:04
  deserves
• 17:04
  the
• 17:04
  skepticism
• 17:05
  the
• 17:05
  community
• 17:06
  is
• 17:06
  giving
• 17:06
  it.
• 17:07
  A
• 17:07
  new
• 17:07
  tool,
• 17:08
  inspired
• 17:08
  by
• 17:08
  the
• 17:09
  CMP170HX
• 17:12
  unlock,
• 17:12
  reportedly
• 17:13
  re-enables
• 17:13
  full-speed
• 17:14
  tensor
• 17:14
  cores
• 17:14
  on
• 17:15
  RTX
• 17:15
  30-
• 17:15
  and
• 17:16
  40-series
• 17:16
  cards,
• 17:17
  with
• 17:17
  one
• 17:17
  tester
• 17:18
  posting
• 17:18
  a
• 17:18
  58%
• 17:19
  BF16
• 17:20
  gain
• 17:20
  on
• 17:20
  dual
• 17:21
  3090s
• 17:21
  and
• 17:22
  headlines
• 17:22
  promising
• 17:23
  up
• 17:23
  to
• 17:23
  2x
• 17:23
  prefill
• 17:24
  .
• 17:24
  The
• 17:25
  sharpest
• 17:25
  reply
• 17:25
  is
• 17:25
  the
• 17:26
  one
• 17:26
  to
• 17:26
  believe:
• 17:26
  most
• 17:27
  backends
• 17:27
  already
• 17:27
  use
• 17:28
  the
• 17:28
  FP16
• 17:29
  path
• 17:29
  that
• 17:29
  sees
• 17:29
  little
• 17:30
  benefit,
• 17:30
  inference
• 17:31
  gains
• 17:31
  land
• 17:32
  around
• 17:32
  10
• 17:32
  to
• 17:32
  15%
• 17:33
  on
• 17:33
  prefill
• 17:33
  with
• 17:34
  no
• 17:34
  change
• 17:34
  to
• 17:34
  decode,
• 17:35
  and
• 17:35
  one
• 17:35
  poster
• 17:36
  argued
• 17:36
  the
• 17:36
  real
• 17:37
  speedup
• 17:37
  someone
• 17:37
  reported
• 17:38
  came
• 17:38
  from
• 17:38
  Claude
• 17:39
  fixing
• 17:39
  a
• 17:39
  broken
• 17:40
  vLLM
• 17:40
  build,
• 17:41
  with
• 17:41
  the
• 17:41
  unlock
• 17:41
  playing
• 17:42
  a
• 17:42
  minor
• 17:42
  role.
• 17:43
  The
• 17:43
  research
• 17:43
  end
• 17:43
  is
• 17:43
  quieter
• 17:44
  and
• 17:44
  more
• 17:44
  interesting.
• 17:45
  MirroS-Lab's
• 17:46
  AgentGarten
• 17:47
  builds
• 17:47
  real-time
• 17:48
  interactive
• 17:48
  worlds
• 17:49
  for
• 17:49
  training
• 17:49
  agents,
• 17:50
  pairing
• 17:50
  program-defined
• 17:51
  simulators
• 17:52
  with
• 17:52
  a
• 17:52
  shared
• 17:52
  neural
• 17:53
  renderer
• 17:53
  adapted
• 17:54
  from
• 17:54
  Cosmos3-Nano,
• 17:56
  so
• 17:56
  agents
• 17:56
  distill
• 17:57
  each
• 17:57
  round
• 17:57
  into
• 17:57
  playbooks
• 17:58
  that
• 17:58
  later
• 17:58
  agents
• 17:59
  inherit
• 17:59
  .
• 18:00
  A
• 18:00
  smaller
• 18:00
  demonstration
• 18:01
  of
• 18:01
  local
• 18:02
  inference
• 18:02
  comes
• 18:02
  from
• 18:03
  a
• 18:03
  browser
• 18:03
  rhythm
• 18:03
  game
• 18:04
  that
• 18:04
  runs
• 18:04
  four
• 18:05
  small
• 18:05
  neural
• 18:05
  networks,
• 18:06
  beat
• 18:06
  tracking,
• 18:07
  note
• 18:07
  transcription,
• 18:08
  vocal
• 18:09
  separation,
• 18:10
  and
• 18:10
  pitch
• 18:10
  tracking,
• 18:11
  about
• 18:11
  22MB
• 18:12
  total,
• 18:13
  entirely
• 18:13
  in
• 18:14
  the
• 18:14
  tab
• 18:14
  on
• 18:14
  onnxruntime-web
• 18:16
  so
• 18:16
  no
• 18:16
  audio
• 18:17
  is
• 18:17
  ever
• 18:17
  uploaded
• 18:17
  .
• 18:18
  And
• 18:18
  for
• 18:18
  anyone
• 18:19
  who
• 18:19
  wants
• 18:19
  the
• 18:19
  device
• 18:20
  itself
• 18:20
  to
• 18:20
  be
• 18:20
  the
• 18:21
  frontier,
• 18:21
  Omarchy
• 18:22
  Mobile
• 18:22
  now
• 18:22
  runs
• 18:23
  native
• 18:23
  Arch
• 18:23
  Linux
• 18:24
  ARM
• 18:24
  with
• 18:24
  Hyprland
• 18:25
  on
• 18:25
  a
• 18:25
  Pixel
• 18:25
  7
• 18:26
  Pro
• 18:26
  with
• 18:26
  no
• 18:26
  Android
• 18:27
  underneath,
• 18:28
  writing
• 18:28
  only
• 18:29
  to
• 18:29
  the
• 18:29
  phone's
• 18:29
  second
• 18:29
  boot
• 18:30
  slot
• 18:30
  so
• 18:30
  stock
• 18:31
  firmware
• 18:31
  stays
• 18:31
  as
• 18:31
  the
• 18:32
  way
• 18:32
  back,
• 18:33
  and
• 18:33
  dropping
• 18:33
  the
• 18:33
  user
• 18:33
  into
• 18:34
  a
• 18:34
  terminal
• 18:34
  running
• 18:35
  an
• 18:35
  AI
• 18:35
  agent
• 18:35
  of
• 18:35
  their
• 18:36
  choice
• 18:36
  at
• 18:36
  first
• 18:36
  login
• 18:37
  .