Latest / AGI Dreams – Open, Uncensored, & Local - AI News Digest / AGI Dreams Podcast – February 20, 2026

Transcript

AGI Dreams Podcast – February 20, 2026 · AGI Dreams – Open, Uncensored, & Local - AI News Digest

• 0:00
  For
• 0:00
  fifteen
• 0:00
  years,
• 0:01
  NVIDIA
• 0:01
  maintained
• 0:02
  a
• 0:02
  clean
• 0:02
  market
• 0:03
  segmentation
• 0:03
  between
• 0:04
  consumer
• 0:04
  and
• 0:04
  enterprise
• 0:05
  GPUs
• 0:05
  using
• 0:06
  a
• 0:06
  single
• 0:06
  lever:
• 0:07
  double-precision
• 0:08
  floating-point.
• 0:09
  On
• 0:09
  the
• 0:09
  Fermi
• 0:09
  architecture
• 0:10
  in
• 0:10
  2010,
• 0:11
  GeForce
• 0:11
  cards
• 0:12
  were
• 0:12
  driver-capped
• 0:12
  to
• 0:12
  a
• 0:12
  1:8
• 0:14
  FP64:FP32
• 0:16
  ratio
• 0:16
  while
• 0:17
  Tesla
• 0:17
  units
• 0:17
  kept
• 0:18
  1:2.
• 0:19
  By
• 0:19
  Ampere
• 0:19
  in
• 0:20
  2020,
• 0:20
  the
• 0:21
  consumer
• 0:21
  ratio
• 0:21
  had
• 0:22
  widened
• 0:22
  to
• 0:22
  1:64
• 0:23
  —
• 0:24
  so
• 0:24
  thoroughly
• 0:24
  that
• 0:24
  NVIDIA's
• 0:25
  own
• 0:25
  GA102
• 0:27
  whitepaper
• 0:27
  described
• 0:28
  the
• 0:28
  residual
• 0:28
  FP64
• 0:29
  units
• 0:30
  as
• 0:30
  existing
• 0:31
  merely
• 0:31
  to
• 0:31
  "ensure
• 0:32
  any
• 0:32
  programs
• 0:32
  with
• 0:32
  FP64
• 0:33
  code
• 0:34
  operate
• 0:34
  correctly."
• 0:35
  A
• 0:35
  detailed
• 0:36
  analysis
• 0:36
  of
• 0:36
  this
• 0:37
  fifteen-year
• 0:37
  degradation
• 0:38
  curve
• 0:38
  finds
• 0:39
  that
• 0:39
  FP64
• 0:40
  throughput
• 0:40
  on
• 0:40
  consumer
• 0:41
  GPUs
• 0:41
  increased
• 0:42
  just
• 0:42
  9.65x
• 0:44
  between
• 0:44
  2010
• 0:44
  and
• 0:44
  2025
• 0:45
  (0.17
• 0:46
  to
• 0:46
  1.64
• 0:48
  TFLOPS),
• 0:50
  while
• 0:50
  FP32
• 0:51
  surged
• 0:51
  77.63x.
• 0:54
  The
• 0:54
  enterprise-to-consumer
• 0:55
  price
• 0:55
  gap
• 0:56
  widened
• 0:56
  from
• 0:56
  roughly
• 0:56
  5x
• 0:57
  in
• 0:57
  2010
• 0:58
  to
• 0:58
  over
• 0:58
  20x
• 0:58
  by
• 0:58
  2022,
• 1:00
  with
• 1:00
  FP64
• 1:01
  as
• 1:01
  the
• 1:01
  convenient
• 1:02
  justification.
• 1:04
  Now
• 1:04
  the
• 1:04
  AI
• 1:04
  boom
• 1:05
  is
• 1:05
  dismantling
• 1:05
  that
• 1:05
  logic.
• 1:07
  Modern
• 1:07
  training
• 1:07
  lives
• 1:08
  comfortably
• 1:08
  in
• 1:08
  FP16,
• 1:09
  BF16,
• 1:10
  FP8,
• 1:11
  and
• 1:11
  even
• 1:12
  FP4
• 1:12
  —
• 1:13
  precisions
• 1:14
  where
• 1:14
  consumer
• 1:14
  GPUs
• 1:15
  are
• 1:15
  surprisingly
• 1:16
  capable.
• 1:17
  The
• 1:17
  Blackwell
• 1:17
  Ultra
• 1:18
  B300
• 1:18
  takes
• 1:19
  the
• 1:19
  reversal
• 1:19
  to
• 1:19
  its
• 1:19
  logical
• 1:20
  conclusion:
• 1:21
  FP64:FP32
• 1:26
  drops
• 1:26
  from
• 1:26
  the
• 1:26
  B200's
• 1:27
  1:2
• 1:28
  to
• 1:28
  the
• 1:28
  B300's
• 1:28
  1:64
• 1:29
  —
• 1:30
  the
• 1:30
  same
• 1:31
  ratio
• 1:31
  as
• 1:31
  a
• 1:31
  consumer
• 1:32
  RTX
• 1:32
  5090
• 1:33
  —
• 1:34
  as
• 1:35
  dedicated
• 1:35
  FP64
• 1:36
  silicon
• 1:37
  is
• 1:37
  cannibalized
• 1:37
  for
• 1:37
  NVFP4
• 1:39
  tensor
• 1:39
  cores.
• 1:41
  Peak
• 1:41
  FP64
• 1:41
  collapses
• 1:42
  from
• 1:42
  37
• 1:43
  TFLOPS
• 1:45
  on
• 1:45
  B200
• 1:46
  to
• 1:46
  1.2
• 1:47
  TFLOPS
• 1:48
  on
• 1:48
  B300.
• 1:50
  NVIDIA
• 1:50
  is
• 1:50
  betting
• 1:51
  on
• 1:51
  FP64
• 1:52
  emulation
• 1:52
  via
• 1:53
  the
• 1:53
  Ozaki
• 1:53
  scheme,
• 1:54
  which
• 1:54
  splits
• 1:54
  double-precision
• 1:55
  numbers
• 1:55
  into
• 1:55
  FP8
• 1:56
  slices,
• 1:57
  multiplies
• 1:57
  them
• 1:57
  on
• 1:58
  tensor
• 1:58
  cores,
• 1:58
  and
• 1:59
  reassembles
• 1:59
  the
• 1:59
  result
• 2:00
  in
• 2:00
  FP64.
• 2:01
  The
• 2:02
  next
• 2:02
  segmentation
• 2:02
  line
• 2:03
  is
• 2:03
  likely
• 2:03
  to
• 2:03
  shift
• 2:03
  from
• 2:04
  FP64
• 2:05
  to
• 2:05
  low-precision
• 2:05
  tensor
• 2:06
  throughput,
• 2:07
  where
• 2:07
  B200
• 2:07
  offers
• 2:08
  a
• 2:08
  16:1
• 2:10
  FP16:FP32
• 2:12
  ratio
• 2:13
  versus
• 2:13
  the
• 2:13
  RTX
• 2:14
  5090's
• 2:15
  1:1.
• 2:16
  The
• 2:16
  capital
• 2:16
  commitments
• 2:17
  are
• 2:17
  recalibrating
• 2:18
  too.
• 2:18
  NVIDIA
• 2:19
  and
• 2:19
  OpenAI
• 2:20
  reportedly
• 2:20
  walked
• 2:21
  away
• 2:21
  from
• 2:21
  an
• 2:21
  unfinished
• 2:22
  $100
• 2:22
  billion
• 2:23
  joint
• 2:23
  venture,
• 2:24
  settling
• 2:25
  instead
• 2:25
  on
• 2:25
  a
• 2:25
  $30
• 2:25
  billion
• 2:26
  investment
• 2:26
  —
• 2:27
  a
• 2:27
  signal
• 2:28
  that
• 2:28
  even
• 2:28
  the
• 2:28
  most
• 2:28
  capital-flush
• 2:29
  players
• 2:29
  in
• 2:30
  AI
• 2:30
  are
• 2:30
  tightening
• 2:30
  scope
• 2:31
  rather
• 2:31
  than
• 2:31
  doubling
• 2:32
  down
• 2:32
  indefinitely.
• 2:34
  Meanwhile,
• 2:34
  Google
• 2:35
  released
• 2:35
  Gemini
• 2:35
  3.1
• 2:36
  Pro
• 2:36
  with
• 2:37
  updated
• 2:37
  benchmarks,
• 2:38
  maintaining
• 2:39
  the
• 2:39
  relentless
• 2:39
  cadence
• 2:40
  of
• 2:40
  frontier
• 2:40
  model
• 2:40
  drops
• 2:41
  that
• 2:41
  keeps
• 2:41
  hardware
• 2:42
  refresh
• 2:42
  pressure
• 2:42
  high.
• 2:45
  The
• 2:45
  team
• 2:45
  behind
• 2:46
  llama.cpp
• 2:46
  —
• 2:47
  the
• 2:47
  C/C++
• 2:48
  library
• 2:49
  that
• 2:49
  made
• 2:49
  local
• 2:49
  model
• 2:50
  inference
• 2:50
  practical
• 2:50
  for
• 2:51
  millions
• 2:51
  —
• 2:51
  is
• 2:52
  joining
• 2:52
  Hugging
• 2:52
  Face.
• 2:53
  GGML
• 2:54
  creator
• 2:54
  Georgi
• 2:55
  Gerganov
• 2:55
  and
• 2:55
  his
• 2:55
  team
• 2:56
  retain
• 2:56
  full
• 2:56
  technical
• 2:57
  autonomy;
• 2:57
  HF
• 2:58
  provides
• 2:58
  long-term
• 2:59
  institutional
• 2:59
  resources.
• 3:00
  The
• 3:00
  concrete
• 3:01
  deliverable:
• 3:02
  near
• 3:02
  single-click
• 3:02
  integration
• 3:03
  between
• 3:03
  HF's
• 3:04
  `transformers`
• 3:04
  library
• 3:05
  (the
• 3:05
  source
• 3:05
  of
• 3:05
  truth
• 3:06
  for
• 3:06
  model
• 3:06
  definitions)
• 3:07
  and
• 3:07
  llama.cpp's
• 3:08
  inference
• 3:09
  engine.
• 3:10
  As
• 3:10
  local
• 3:10
  inference
• 3:10
  becomes
• 3:11
  a
• 3:11
  competitive
• 3:11
  alternative
• 3:12
  to
• 3:12
  cloud
• 3:12
  endpoints,
• 3:13
  the
• 3:13
  packaging
• 3:13
  gap
• 3:14
  matters
• 3:14
  more
• 3:14
  than
• 3:14
  raw
• 3:15
  capability.
• 3:16
  HF
• 3:16
  is
• 3:16
  positioning
• 3:17
  llama.cpp
• 3:18
  as
• 3:19
  the
• 3:19
  ubiquitous
• 3:19
  local
• 3:19
  inference
• 3:20
  layer
• 3:20
  that
• 3:20
  "runs
• 3:20
  as
• 3:21
  efficiently
• 3:21
  as
• 3:21
  possible
• 3:21
  on
• 3:22
  our
• 3:22
  devices,"
• 3:23
  with
• 3:23
  shared
• 3:23
  ambitions
• 3:23
  to
• 3:24
  "make
• 3:24
  open-source
• 3:24
  superintelligence
• 3:25
  accessible
• 3:25
  to
• 3:25
  the
• 3:26
  world."
• 3:27
  One
• 3:27
  enterprise
• 3:27
  team
• 3:27
  demonstrates
• 3:28
  the
• 3:28
  practical
• 3:28
  endgame.
• 3:30
  Two
• 3:30
  NVIDIA
• 3:30
  DGX
• 3:31
  Spark
• 3:31
  units
• 3:31
  ($8,000
• 3:32
  total)
• 3:32
  deliver
• 3:32
  256GB
• 3:34
  unified
• 3:34
  memory
• 3:35
  and
• 3:35
  2
• 3:35
  PFLOPS,
• 3:36
  running
• 3:37
  Alibaba's
• 3:37
  Qwen3-Coder-Next
• 3:37
  —
• 3:39
  an
• 3:39
  80B-parameter
• 3:39
  MoE
• 3:40
  model
• 3:40
  with
• 3:40
  only
• 3:40
  3B
• 3:41
  active
• 3:41
  parameters
• 3:41
  per
• 3:41
  inference
• 3:42
  and
• 3:42
  256K
• 3:43
  native
• 3:43
  context
• 3:44
  —
• 3:44
  at
• 3:45
  Q8
• 3:45
  quantization
• 3:45
  with
• 3:45
  zero
• 3:46
  API
• 3:46
  costs.
• 3:48
  The
• 3:48
  team
• 3:48
  deploys
• 3:48
  it
• 3:48
  as
• 3:48
  an
• 3:49
  agentic
• 3:49
  coding
• 3:49
  assistant
• 3:50
  that
• 3:50
  implements
• 3:50
  across
• 3:50
  multiple
• 3:51
  files,
• 3:52
  delivers
• 3:52
  PRs
• 3:53
  with
• 3:53
  blast
• 3:53
  radius
• 3:53
  analysis,
• 3:54
  and
• 3:54
  runs
• 3:55
  a
• 3:55
  "shadow
• 3:55
  engineering
• 3:55
  agent"
• 3:56
  that
• 3:56
  watches
• 3:57
  human
• 3:57
  PRs
• 3:57
  and
• 3:58
  flags
• 3:58
  quality
• 3:58
  issues
• 3:59
  continuously.
• 4:00
  Agent
• 4:00
  orchestration
• 4:01
  remains
• 4:01
  harder
• 4:01
  than
• 4:01
  expected:
• 4:02
  the
• 4:02
  model
• 4:02
  occasionally
• 4:03
  hallucinates
• 4:03
  file
• 4:04
  paths
• 4:04
  or
• 4:04
  refactors
• 4:05
  signatures
• 4:05
  without
• 4:05
  updating
• 4:06
  all
• 4:06
  call
• 4:06
  sites,
• 4:07
  requiring
• 4:07
  validation
• 4:08
  loops
• 4:08
  the
• 4:08
  agent
• 4:09
  runs
• 4:09
  against
• 4:09
  its
• 4:09
  own
• 4:09
  output.
• 4:10
  Commenters
• 4:11
  noted
• 4:11
  the
• 4:11
  $8,000
• 4:12
  represents
• 4:13
  40
• 4:13
  months
• 4:13
  of
• 4:13
  a
• 4:13
  Claude
• 4:14
  Max
• 4:14
  subscription
• 4:14
  —
• 4:15
  the
• 4:15
  ROI
• 4:15
  depends
• 4:16
  heavily
• 4:16
  on
• 4:16
  throughput,
• 4:16
  concurrency,
• 4:17
  and
• 4:17
  data
• 4:17
  sovereignty
• 4:18
  requirements.
• 4:19
  The
• 4:19
  sovereignty
• 4:20
  angle
• 4:20
  matters
• 4:20
  more
• 4:20
  than
• 4:21
  the
• 4:21
  team
• 4:21
  expected:
• 4:22
  every
• 4:22
  prompt,
• 4:23
  every
• 4:23
  model
• 4:23
  weight
• 4:24
  stays
• 4:24
  in-building.
• 4:25
  The
• 4:25
  tooling
• 4:26
  ecosystem
• 4:26
  continues
• 4:27
  broadening
• 4:27
  beyond
• 4:27
  developer
• 4:28
  terminals:
• 4:29
  Vellium,
• 4:30
  a
• 4:30
  new
• 4:30
  open-source
• 4:31
  desktop
• 4:31
  app,
• 4:32
  brings
• 4:32
  visual
• 4:32
  controls
• 4:33
  for
• 4:33
  creative
• 4:33
  writing
• 4:33
  with
• 4:33
  local
• 4:34
  models
• 4:34
  to
• 4:34
  non-technical
• 4:35
  users.
• 4:38
  Some
• 4:38
  transformer
• 4:38
  attention
• 4:39
  heads
• 4:39
  appear
• 4:39
  to
• 4:39
  function
• 4:40
  as
• 4:40
  membership
• 4:40
  testers
• 4:40
  —
• 4:41
  dedicating
• 4:42
  themselves
• 4:42
  to
• 4:42
  answering
• 4:43
  "has
• 4:43
  this
• 4:43
  token
• 4:44
  appeared
• 4:44
  before
• 4:44
  in
• 4:44
  the
• 4:44
  context?"
• 4:46
  A
• 4:46
  new
• 4:46
  interpretability
• 4:47
  paper
• 4:47
  identifies
• 4:48
  three
• 4:48
  such
• 4:48
  heads
• 4:48
  in
• 4:49
  GPT-2
• 4:49
  small
• 4:50
  (L0H1,
• 4:51
  L0H5,
• 4:53
  L1H11)
• 4:55
  that
• 4:55
  form
• 4:55
  a
• 4:55
  multi-resolution
• 4:56
  membership-testing
• 4:57
  system
• 4:57
  concentrated
• 4:57
  in
• 4:58
  layers
• 4:58
  0–1,
• 5:00
  taxonomically
• 5:00
  distinct
• 5:01
  from
• 5:01
  both
• 5:01
  induction
• 5:01
  heads
• 5:02
  and
• 5:02
  previous-token
• 5:03
  heads.
• 5:04
  The
• 5:04
  evidence
• 5:04
  is
• 5:04
  precise:
• 5:05
  head
• 5:05
  L1H11
• 5:07
  follows
• 5:08
  the
• 5:08
  theoretical
• 5:08
  Bloom
• 5:09
  filter
• 5:09
  capacity
• 5:09
  formula
• 5:10
  (k=1
• 5:11
  hash
• 5:11
  function,
• 5:12
  ~22
• 5:13
  fitted
• 5:13
  capacity
• 5:13
  bits),
• 5:14
  saturating
• 5:14
  by
• 5:15
  ~40
• 5:15
  unique
• 5:15
  context
• 5:16
  tokens.
• 5:17
  The
• 5:17
  other
• 5:17
  two
• 5:17
  heads
• 5:17
  are
• 5:17
  more
• 5:18
  interesting
• 5:18
  —
• 5:18
  they
• 5:18
  maintain
• 5:19
  false
• 5:19
  positive
• 5:19
  rates
• 5:20
  of
• 5:20
  0–4%
• 5:21
  even
• 5:22
  at
• 5:22
  180
• 5:22
  unique
• 5:23
  tokens,
• 5:23
  well
• 5:24
  exceeding
• 5:24
  classical
• 5:24
  Bloom
• 5:25
  filter
• 5:25
  bounds
• 5:25
  and
• 5:26
  suggesting
• 5:26
  a
• 5:26
  richer
• 5:26
  mechanism
• 5:27
  than
• 5:27
  simple
• 5:28
  bit-array
• 5:28
  hashing.
• 5:29
  A
• 5:29
  similarity
• 5:30
  sweep
• 5:30
  across
• 5:30
  1,284
• 5:32
  controlled
• 5:33
  probe
• 5:33
  tokens
• 5:34
  reveals
• 5:34
  that
• 5:34
  false
• 5:34
  positive
• 5:35
  rates
• 5:35
  decay
• 5:35
  monotonically
• 5:36
  with
• 5:36
  cosine
• 5:37
  distance
• 5:37
  in
• 5:37
  the
• 5:38
  embedding
• 5:38
  space
• 5:38
  —
• 5:39
  consistent
• 5:40
  with
• 5:40
  locality-sensitive
• 5:41
  hashing,
• 5:41
  not
• 5:41
  uniform
• 5:42
  hashing.
• 5:43
  The
• 5:43
  "hash
• 5:43
  functions"
• 5:44
  are
• 5:44
  the
• 5:44
  QK
• 5:44
  projections
• 5:45
  themselves.
• 5:46
  L0H5
• 5:47
  (ultra-precise)
• 5:48
  reaches
• 5:48
  noise
• 5:48
  floor
• 5:49
  by
• 5:49
  cosine
• 5:49
  0.5;
• 5:50
  L1H11
• 5:52
  (broad)
• 5:53
  retains
• 5:53
  17%
• 5:54
  false
• 5:55
  positive
• 5:55
  rate
• 5:55
  at
• 5:55
  that
• 5:55
  threshold.
• 5:57
  Together
• 5:57
  the
• 5:57
  three
• 5:57
  heads
• 5:58
  give
• 5:58
  the
• 5:58
  model
• 5:58
  membership
• 5:59
  signals
• 5:59
  at
• 5:59
  multiple
• 6:00
  similarity
• 6:00
  resolutions,
• 6:01
  paralleling
• 6:02
  the
• 6:02
  multi-granularity
• 6:03
  Bloom
• 6:03
  filter
• 6:04
  designs
• 6:04
  human
• 6:05
  researchers
• 6:05
  proposed
• 6:06
  deliberately.
• 6:07
  A
• 6:07
  fourth
• 6:07
  candidate
• 6:08
  (L3H0)
• 6:09
  was
• 6:09
  reclassified
• 6:09
  as
• 6:10
  a
• 6:10
  generic
• 6:10
  prefix-attention
• 6:11
  head
• 6:11
  after
• 6:11
  confound
• 6:12
  controls
• 6:12
  revealed
• 6:13
  its
• 6:13
  capacity
• 6:13
  curve
• 6:14
  was
• 6:14
  a
• 6:14
  sequence-length
• 6:14
  artifact
• 6:15
  —
• 6:15
  a
• 6:15
  self-correction
• 6:16
  the
• 6:16
  authors
• 6:17
  argue
• 6:17
  strengthens
• 6:17
  the
• 6:17
  surviving
• 6:18
  findings.
• 6:19
  The
• 6:19
  practical
• 6:19
  implication
• 6:20
  is
• 6:20
  immediate:
• 6:21
  because
• 6:21
  these
• 6:21
  heads
• 6:22
  are
• 6:22
  already
• 6:22
  computed
• 6:22
  during
• 6:23
  inference,
• 6:24
  monitoring
• 6:24
  their
• 6:24
  false
• 6:25
  positive
• 6:25
  activations
• 6:26
  provides
• 6:26
  a
• 6:26
  zero-cost
• 6:27
  hallucination
• 6:27
  diagnostic
• 6:28
  at
• 6:28
  the
• 6:28
  layer
• 6:29
  where
• 6:29
  the
• 6:29
  error
• 6:29
  originates.
• 6:31
  The
• 6:31
  deeper
• 6:31
  scientific
• 6:31
  point
• 6:32
  is
• 6:32
  striking:
• 6:33
  gradient
• 6:33
  descent,
• 6:34
  optimizing
• 6:35
  nothing
• 6:35
  but
• 6:35
  next-token
• 6:36
  prediction,
• 6:37
  converges
• 6:37
  independently
• 6:38
  on
• 6:38
  a
• 6:38
  variant
• 6:39
  of
• 6:39
  the
• 6:39
  distance-sensitive
• 6:40
  Bloom
• 6:40
  filter
• 6:40
  that
• 6:41
  human
• 6:41
  researchers
• 6:42
  took
• 6:42
  36
• 6:42
  years
• 6:43
  to
• 6:43
  formalize.
• 6:44
  Ironically,
• 6:45
  throwing
• 6:45
  more
• 6:46
  compute
• 6:46
  at
• 6:46
  inference
• 6:47
  doesn't
• 6:47
  reliably
• 6:47
  help.
• 6:48
  An
• 6:48
  evaluation
• 6:49
  of
• 6:49
  22
• 6:49
  model
• 6:50
  configurations
• 6:51
  on
• 6:51
  Deep
• 6:51
  Research
• 6:51
  Bench
• 6:52
  (169
• 6:53
  web-research
• 6:53
  tasks,
• 6:54
  human-verified
• 6:55
  answers)
• 6:56
  found
• 6:56
  that
• 6:56
  GPT-5
• 6:57
  at
• 6:57
  high
• 6:57
  effort
• 6:57
  scored
• 6:58
  0.481
• 6:59
  versus
• 7:00
  0.496
• 7:01
  at
• 7:01
  low
• 7:01
  effort
• 7:02
  —
• 7:02
  while
• 7:03
  costing
• 7:03
  55%
• 7:04
  more
• 7:04
  per
• 7:04
  query
• 7:05
  ($0.39
• 7:06
  vs
• 7:07
  $0.25).
• 7:07
  Gemini-3-Flash
• 7:08
  showed
• 7:09
  a
• 7:09
  5-point
• 7:09
  drop
• 7:09
  from
• 7:10
  low
• 7:10
  to
• 7:10
  high
• 7:10
  effort.
• 7:11
  The
• 7:11
  researchers'
• 7:12
  takeaway:
• 7:13
  minimal
• 7:13
  effort
• 7:13
  settings
• 7:14
  often
• 7:14
  occupy
• 7:15
  the
• 7:15
  Pareto
• 7:15
  frontier
• 7:16
  for
• 7:16
  accuracy-per-dollar.
• 7:19
  IBM
• 7:20
  Research
• 7:20
  and
• 7:21
  UC
• 7:21
  Berkeley
• 7:21
  applied
• 7:21
  MAST
• 7:22
  (Multi-Agent
• 7:22
  System
• 7:22
  Failure
• 7:23
  Taxonomy)
• 7:23
  to
• 7:23
  ITBench
• 7:24
  —
• 7:25
  the
• 7:25
  industry
• 7:25
  benchmark
• 7:25
  for
• 7:25
  SRE,
• 7:26
  security,
• 7:26
  and
• 7:26
  FinOps
• 7:27
  automation
• 7:27
  —
• 7:28
  annotating
• 7:28
  310
• 7:29
  execution
• 7:29
  traces
• 7:30
  across
• 7:30
  three
• 7:30
  model
• 7:30
  tiers.
• 7:31
  The
• 7:31
  results
• 7:32
  map
• 7:32
  the
• 7:32
  failure
• 7:32
  landscape
• 7:33
  with
• 7:33
  unusual
• 7:33
  precision.
• 7:34
  Frontier
• 7:35
  models
• 7:35
  like
• 7:35
  Gemini-3-Flash
• 7:36
  fail
• 7:36
  surgically:
• 7:37
  2.6
• 7:38
  failure
• 7:38
  modes
• 7:39
  per
• 7:39
  trace,
• 7:40
  typically
• 7:40
  hitting
• 7:40
  an
• 7:41
  isolated
• 7:41
  bottleneck.
• 7:42
  Large
• 7:43
  open
• 7:43
  models
• 7:43
  like
• 7:43
  GPT-OSS-120B
• 7:48
  suffer
• 7:49
  cascading
• 7:49
  collapse:
• 7:50
  5.3
• 7:51
  failure
• 7:51
  modes
• 7:51
  per
• 7:52
  trace,
• 7:52
  where
• 7:53
  a
• 7:53
  single
• 7:53
  reasoning
• 7:53
  mismatch
• 7:54
  early
• 7:54
  in
• 7:54
  a
• 7:54
  run
• 7:54
  poisons
• 7:55
  the
• 7:55
  context
• 7:55
  and
• 7:55
  compounds
• 7:56
  into
• 7:56
  total
• 7:56
  derailment.
• 7:57
  The
• 7:57
  taxonomy
• 7:58
  identifies
• 7:59
  14
• 7:59
  distinct
• 7:59
  failure
• 8:00
  patterns
• 8:00
  across
• 8:00
  structural
• 8:01
  (looping,
• 8:02
  memory
• 8:02
  loss,
• 8:02
  premature
• 8:03
  termination),
• 8:04
  communication
• 8:04
  (inability
• 8:05
  to
• 8:05
  clarify,
• 8:06
  going
• 8:06
  off-topic),
• 8:07
  and
• 8:07
  quality
• 8:07
  control
• 8:07
  categories
• 8:08
  (incorrect
• 8:08
  verification,
• 8:09
  hallucinated
• 8:10
  success).
• 8:11
  The
• 8:11
  single
• 8:11
  strongest
• 8:12
  predictor
• 8:12
  of
• 8:12
  failure
• 8:13
  is
• 8:13
  FM-3.3:
• 8:14
  Incorrect
• 8:15
  Verification
• 8:15
  —
• 8:16
  agents
• 8:16
  "declaring
• 8:17
  victory"
• 8:17
  without
• 8:18
  checking
• 8:18
  ground
• 8:18
  truth.
• 8:19
  For
• 8:20
  Gemini-3-Flash,
• 8:21
  verification
• 8:22
  errors
• 8:22
  spike
• 8:22
  52%
• 8:23
  in
• 8:23
  failed
• 8:23
  traces
• 8:24
  versus
• 8:24
  successful
• 8:25
  ones.
• 8:26
  Kimi-K2
• 8:26
  has
• 8:27
  a
• 8:27
  different
• 8:27
  pathology:
• 8:28
  a
• 8:28
  +46%
• 8:29
  premature
• 8:29
  termination
• 8:30
  rate
• 8:30
  and
• 8:31
  a
• 8:31
  staggering
• 8:31
  92%
• 8:32
  prevalence
• 8:32
  of
• 8:32
  reasoning-action
• 8:33
  mismatch
• 8:33
  in
• 8:34
  failed
• 8:34
  runs,
• 8:35
  where
• 8:35
  the
• 8:35
  model
• 8:35
  identifies
• 8:36
  the
• 8:36
  correct
• 8:36
  next
• 8:36
  step
• 8:37
  then
• 8:37
  executes
• 8:37
  an
• 8:37
  irrelevant
• 8:38
  command.
• 8:39
  GPT-OSS-120B
• 8:43
  loses
• 8:44
  conversation
• 8:44
  history
• 8:45
  in
• 8:45
  the
• 8:45
  majority
• 8:45
  of
• 8:46
  its
• 8:46
  traces
• 8:46
  (versus
• 8:46
  zero
• 8:47
  for
• 8:47
  Gemini),
• 8:48
  and
• 8:48
  87%
• 8:49
  of
• 8:49
  failed
• 8:49
  runs
• 8:49
  show
• 8:50
  decoupled
• 8:50
  reasoning
• 8:50
  and
• 8:51
  action.
• 8:52
  The
• 8:52
  crucial
• 8:52
  finding:
• 8:53
  prompt
• 8:53
  engineering
• 8:54
  alone
• 8:54
  yields
• 8:55
  ~15%
• 8:55
  improvement;
• 8:57
  architectural
• 8:57
  interventions
• 8:58
  —
• 8:58
  context-management
• 8:59
  agents,
• 9:00
  finite
• 9:00
  state
• 9:01
  machines
• 9:01
  for
• 9:01
  termination
• 9:01
  —
• 9:02
  deliver
• 9:03
  up
• 9:03
  to
• 9:03
  53%.
• 9:04
  Two
• 9:04
  models
• 9:05
  can
• 9:05
  share
• 9:05
  the
• 9:05
  same
• 9:06
  success
• 9:06
  rate
• 9:06
  yet
• 9:06
  fail
• 9:07
  for
• 9:07
  entirely
• 9:07
  different
• 9:08
  structural
• 9:08
  reasons
• 9:08
  requiring
• 9:09
  entirely
• 9:09
  different
• 9:10
  fixes.
• 9:11
  This
• 9:11
  pattern
• 9:11
  surfaces
• 9:12
  independently
• 9:12
  in
• 9:13
  a
• 9:13
  forensic
• 9:13
  audit
• 9:13
  where
• 9:14
  one
• 9:14
  user
• 9:14
  found
• 9:14
  40.8%
• 9:16
  of
• 9:16
  their
• 9:16
  local
• 9:16
  AI
• 9:16
  assistant's
• 9:17
  claimed
• 9:17
  task
• 9:17
  completions
• 9:18
  were
• 9:18
  outright
• 9:18
  fabricated
• 9:19
  —
• 9:20
  a
• 9:20
  blunt
• 9:20
  reminder
• 9:21
  that
• 9:21
  verification
• 9:21
  cannot
• 9:22
  be
• 9:22
  optional.
• 9:23
  OpenAI's
• 9:24
  newly
• 9:24
  published
• 9:25
  "Practical
• 9:25
  Guide
• 9:25
  to
• 9:25
  Building
• 9:26
  Agents"
• 9:26
  arrives
• 9:27
  as
• 9:27
  if
• 9:27
  timed
• 9:27
  to
• 9:27
  this
• 9:28
  diagnosis.
• 9:29
  It
• 9:29
  advocates
• 9:30
  layered
• 9:30
  guardrails
• 9:30
  (relevance
• 9:31
  classifiers,
• 9:32
  safety
• 9:32
  classifiers,
• 9:33
  PII
• 9:33
  filters,
• 9:34
  tool
• 9:35
  risk
• 9:35
  ratings
• 9:35
  operating
• 9:36
  concurrently
• 9:36
  with
• 9:36
  agent
• 9:37
  execution),
• 9:38
  single-agent-first
• 9:39
  design
• 9:39
  with
• 9:39
  multi-agent
• 9:40
  architectures
• 9:41
  reserved
• 9:41
  for
• 9:41
  high-complexity
• 9:42
  workflows,
• 9:43
  and
• 9:43
  explicit
• 9:43
  human
• 9:44
  intervention
• 9:44
  triggers
• 9:44
  for
• 9:45
  high-risk
• 9:45
  or
• 9:45
  irreversible
• 9:46
  actions.
• 9:47
  The
• 9:47
  most
• 9:47
  practical
• 9:48
  recommendation:
• 9:49
  never
• 9:49
  let
• 9:50
  the
• 9:50
  LLM
• 9:50
  grade
• 9:51
  its
• 9:51
  own
• 9:51
  homework.
• 9:54
  If
• 9:54
  incorrect
• 9:55
  verification
• 9:55
  is
• 9:55
  the
• 9:55
  Achilles'
• 9:56
  heel,
• 9:56
  Daniel
• 9:57
  Miessler's
• 9:57
  "Generalized
• 9:58
  Hill
• 9:58
  Climbing
• 9:59
  Runtime"
• 9:59
  proposes
• 10:00
  making
• 10:00
  verification
• 10:01
  the
• 10:01
  entire
• 10:01
  architecture.
• 10:03
  Built
• 10:03
  on
• 10:03
  top
• 10:04
  of
• 10:04
  Claude
• 10:04
  Code,
• 10:04
  the
• 10:04
  framework
• 10:05
  centers
• 10:05
  on
• 10:05
  Ideal
• 10:05
  State
• 10:05
  Criteria
• 10:06
  (ISC)
• 10:07
  —
• 10:07
  8-12
• 10:08
  word,
• 10:08
  binary-testable,
• 10:09
  boolean
• 10:09
  statements
• 10:10
  that
• 10:10
  define
• 10:10
  success
• 10:11
  before
• 10:11
  any
• 10:11
  work
• 10:11
  begins.
• 10:13
  Every
• 10:13
  request
• 10:13
  is
• 10:13
  reverse-engineered:
• 10:15
  What
• 10:15
  was
• 10:15
  explicitly
• 10:15
  asked
• 10:16
  for?
• 10:17
  What
• 10:17
  are
• 10:17
  the
• 10:17
  gotchas?
• 10:18
  What
• 10:19
  common
• 10:19
  failure
• 10:19
  modes
• 10:19
  exist?
• 10:20
  The
• 10:21
  resulting
• 10:21
  ISC
• 10:21
  serve
• 10:22
  as
• 10:22
  both
• 10:22
  specification
• 10:23
  and
• 10:23
  verification
• 10:24
  gates.
• 10:25
  An
• 10:25
  effort-level
• 10:25
  controller
• 10:26
  scales
• 10:26
  the
• 10:26
  machinery
• 10:27
  from
• 10:27
  instant
• 10:27
  fixes
• 10:27
  (4–8
• 10:28
  criteria)
• 10:29
  to
• 10:29
  deep
• 10:29
  system
• 10:30
  redesigns
• 10:30
  (40–150+
• 10:32
  criteria
• 10:33
  with
• 10:33
  parallel
• 10:33
  agent
• 10:33
  spawning
• 10:34
  and
• 10:34
  multi-agent
• 10:34
  debate).
• 10:35
  The
• 10:36
  inner
• 10:36
  loop
• 10:36
  is
• 10:36
  the
• 10:36
  scientific
• 10:37
  method
• 10:37
  —
• 10:38
  observe,
• 10:38
  hypothesize,
• 10:40
  test,
• 10:40
  iterate
• 10:40
  —
• 10:41
  running
• 10:41
  against
• 10:42
  the
• 10:42
  outer
• 10:42
  loop
• 10:42
  of
• 10:42
  current-to-ideal
• 10:43
  state.
• 10:44
  Miessler
• 10:45
  reports
• 10:45
  meaningful
• 10:46
  reduction
• 10:46
  in
• 10:46
  drift
• 10:46
  and
• 10:47
  rework
• 10:47
  when
• 10:48
  ISC
• 10:48
  are
• 10:48
  combined
• 10:48
  with
• 10:49
  execution
• 10:49
  in
• 10:49
  waves,
• 10:50
  with
• 10:50
  deterministic
• 10:51
  gates
• 10:51
  between
• 10:51
  waves
• 10:52
  that
• 10:52
  check
• 10:52
  deliverables
• 10:52
  before
• 10:53
  authorizing
• 10:53
  the
• 10:53
  next
• 10:54
  phase.
• 10:54
  "If
• 10:55
  you
• 10:55
  can't
• 10:55
  articulate
• 10:56
  what
• 10:56
  you
• 10:56
  want,
• 10:56
  prompting
• 10:57
  and
• 10:57
  context
• 10:57
  won't
• 10:58
  help
• 10:58
  you
• 10:58
  much."
• 10:59
  The
• 10:59
  "Build
• 10:59
  with
• 10:59
  Quality"
• 11:00
  skill
• 11:00
  demonstrates
• 11:01
  what
• 11:01
  this
• 11:01
  looks
• 11:01
  like
• 11:01
  at
• 11:01
  scale:
• 11:02
  a
• 11:03
  111-agent
• 11:04
  swarm
• 11:04
  in
• 11:04
  a
• 11:04
  hierarchical-mesh
• 11:05
  topology
• 11:06
  across
• 11:06
  development,
• 11:07
  quality,
• 11:07
  and
• 11:07
  security
• 11:07
  domains.
• 11:09
  Its
• 11:09
  validation
• 11:10
  project
• 11:10
  —
• 11:11
  ShopFlow
• 11:11
  V2,
• 11:12
  a
• 11:12
  4,220-line
• 11:14
  e-commerce
• 11:14
  platform
• 11:15
  with
• 11:15
  6
• 11:15
  bounded
• 11:15
  contexts,
• 11:16
  5
• 11:17
  architecture
• 11:17
  decision
• 11:18
  records,
• 11:18
  and
• 11:18
  90%+
• 11:19
  test
• 11:19
  coverage
• 11:19
  —
• 11:20
  was
• 11:20
  built
• 11:21
  in
• 11:21
  a
• 11:21
  single
• 11:21
  session
• 11:21
  with
• 11:22
  real
• 11:22
  Stripe
• 11:22
  payments
• 11:22
  and
• 11:23
  PostgreSQL
• 11:23
  persistence.
• 11:25
  Agents
• 11:25
  communicate
• 11:25
  laterally:
• 11:26
  the
• 11:26
  test
• 11:27
  generator
• 11:27
  knows
• 11:27
  what
• 11:27
  the
• 11:27
  coder
• 11:28
  built,
• 11:28
  the
• 11:28
  security
• 11:29
  scanner
• 11:29
  feeds
• 11:29
  findings
• 11:30
  to
• 11:30
  the
• 11:30
  architect.
• 11:31
  The
• 11:31
  explicit
• 11:32
  pitch:
• 11:32
  "Can
• 11:33
  AI
• 11:33
  write
• 11:33
  code?
• 11:34
  Yes.
• 11:35
  So
• 11:35
  can
• 11:35
  offshore
• 11:36
  teams.
• 11:37
  The
• 11:37
  question
• 11:37
  is
• 11:37
  whether
• 11:37
  it
• 11:38
  can
• 11:38
  write
• 11:38
  code
• 11:38
  your
• 11:38
  team
• 11:38
  can
• 11:39
  maintain."
• 11:43
  "AI
• 11:43
  slop
• 11:43
  PRs
• 11:44
  are
• 11:44
  becoming
• 11:44
  increasingly
• 11:45
  draining
• 11:45
  and
• 11:45
  demoralizing
• 11:46
  for
• 11:46
  Godot
• 11:46
  maintainers,"
• 11:47
  writes
• 11:48
  Godot
• 11:48
  project
• 11:48
  manager
• 11:49
  Rémi
• 11:49
  Verschelde,
• 11:50
  capturing
• 11:51
  the
• 11:51
  central
• 11:51
  tension
• 11:51
  in
• 11:51
  a
• 11:51
  detailed
• 11:52
  examination
• 11:52
  of
• 11:53
  AI-generated
• 11:53
  code's
• 11:53
  impact
• 11:54
  on
• 11:54
  open-source
• 11:55
  economics.
• 11:56
  The
• 11:56
  problem
• 11:56
  isn't
• 11:57
  bad
• 11:57
  code
• 11:57
  —
• 11:57
  it's
• 11:57
  convincing
• 11:58
  code.
• 11:59
  "It
• 11:59
  compiles,
• 12:00
  it
• 12:00
  passes
• 12:00
  superficial
• 12:01
  review
• 12:01
  and
• 12:01
  it
• 12:02
  looks
• 12:02
  professional,
• 12:03
  but
• 12:03
  it
• 12:03
  may
• 12:03
  embed
• 12:04
  subtle
• 12:04
  logic
• 12:04
  errors,
• 12:05
  security
• 12:06
  flaws,
• 12:06
  or
• 12:07
  unmaintainable
• 12:07
  complexity,"
• 12:08
  notes
• 12:09
  Hiswai
• 12:09
  CTO
• 12:10
  Peter
• 12:10
  Vincalek.
• 12:11
  A
• 12:11
  contributor
• 12:12
  generates
• 12:12
  a
• 12:12
  500-line
• 12:13
  PR
• 12:13
  in
• 12:13
  90
• 12:14
  seconds;
• 12:15
  a
• 12:15
  maintainer
• 12:15
  still
• 12:15
  needs
• 12:16
  2
• 12:16
  hours
• 12:16
  to
• 12:16
  evaluate
• 12:17
  it.
• 12:18
  The
• 12:18
  cost
• 12:18
  of
• 12:18
  producing
• 12:18
  code
• 12:19
  has
• 12:19
  cratered
• 12:19
  to
• 12:19
  near
• 12:19
  zero;
• 12:20
  the
• 12:20
  cost
• 12:21
  of
• 12:21
  reviewing
• 12:21
  it
• 12:21
  hasn't
• 12:22
  budged.
• 12:23
  One
• 12:23
  researcher
• 12:23
  found
• 12:23
  that
• 12:24
  ~20%
• 12:24
  of
• 12:24
  package
• 12:25
  names
• 12:25
  in
• 12:25
  AI-generated
• 12:26
  code
• 12:27
  don't
• 12:27
  even
• 12:27
  exist
• 12:27
  —
• 12:28
  and
• 12:28
  attackers
• 12:29
  are
• 12:29
  already
• 12:29
  squatting
• 12:29
  those
• 12:30
  phantom
• 12:30
  names,
• 12:31
  turning
• 12:31
  hallucination
• 12:32
  into
• 12:32
  a
• 12:32
  supply
• 12:32
  chain
• 12:32
  attack
• 12:33
  vector.
• 12:34
  The
• 12:34
  irony:
• 12:35
  some
• 12:35
  enterprises
• 12:36
  moved
• 12:36
  to
• 12:36
  open
• 12:36
  source
• 12:37
  specifically
• 12:37
  to
• 12:37
  escape
• 12:38
  hyperscaler
• 12:38
  AI
• 12:38
  risk,
• 12:39
  only
• 12:39
  to
• 12:39
  discover
• 12:40
  AI-generated
• 12:41
  code
• 12:41
  now
• 12:41
  flowing
• 12:42
  into
• 12:42
  open
• 12:42
  source
• 12:42
  itself.
• 12:43
  As
• 12:44
  one
• 12:44
  consultant
• 12:44
  put
• 12:44
  it,
• 12:45
  "AI
• 12:45
  slop
• 12:46
  creates
• 12:46
  a
• 12:46
  false
• 12:46
  sense
• 12:47
  of
• 12:47
  velocity.
• 12:48
  You
• 12:48
  think
• 12:48
  you're
• 12:48
  shipping
• 12:49
  faster,
• 12:49
  but
• 12:49
  you're
• 12:50
  actually
• 12:50
  accumulating
• 12:51
  risk
• 12:51
  faster
• 12:51
  than
• 12:52
  your
• 12:52
  team
• 12:52
  can
• 12:52
  pay
• 12:52
  it
• 12:52
  down."
• 12:54
  The
• 12:54
  "Think
• 12:54
  Tax"
• 12:55
  editorial
• 12:55
  frames
• 12:55
  this
• 12:56
  as
• 12:56
  cognitive
• 12:56
  atrophy,
• 12:57
  not
• 12:57
  tooling
• 12:58
  failure.
• 12:59
  Every
• 12:59
  AI-generated
• 13:00
  function
• 13:00
  accepted
• 13:01
  without
• 13:01
  building
• 13:01
  a
• 13:01
  mental
• 13:02
  model
• 13:02
  defers
• 13:02
  a
• 13:02
  payment
• 13:03
  that
• 13:03
  compounds
• 13:03
  exponentially.
• 13:05
  Anthropic's
• 13:05
  own
• 13:06
  research
• 13:06
  found
• 13:06
  developers
• 13:07
  using
• 13:07
  AI
• 13:08
  scored
• 13:08
  17%
• 13:09
  lower
• 13:09
  on
• 13:09
  comprehension
• 13:10
  when
• 13:10
  learning
• 13:10
  new
• 13:11
  libraries.
• 13:12
  Stack
• 13:12
  Overflow's
• 13:13
  2025
• 13:13
  survey
• 13:14
  puts
• 13:14
  AI
• 13:14
  tool
• 13:15
  usage
• 13:15
  at
• 13:15
  85%
• 13:16
  of
• 13:16
  developers
• 13:16
  —
• 13:17
  a
• 13:17
  systemic
• 13:18
  comprehension
• 13:18
  gap
• 13:18
  forming
• 13:19
  across
• 13:19
  the
• 13:19
  industry.
• 13:20
  The
• 13:20
  progression
• 13:21
  is
• 13:21
  predictable:
• 13:22
  a
• 13:22
  honeymoon
• 13:23
  where
• 13:23
  AI
• 13:23
  delivers
• 13:24
  genuine
• 13:24
  productivity
• 13:25
  gains,
• 13:26
  a
• 13:26
  drift
• 13:26
  where
• 13:26
  "probably
• 13:27
  won't
• 13:27
  break"
• 13:28
  becomes
• 13:28
  the
• 13:28
  approval
• 13:28
  standard,
• 13:29
  and
• 13:30
  a
• 13:30
  cliff
• 13:30
  where
• 13:30
  debugging
• 13:30
  becomes
• 13:31
  archaeology.
• 13:32
  The
• 13:32
  analogy
• 13:33
  is
• 13:33
  uncomfortable:
• 13:34
  GPS
• 13:35
  killed
• 13:35
  spatial
• 13:35
  awareness,
• 13:36
  autocorrect
• 13:37
  degraded
• 13:37
  spelling,
• 13:38
  autopilot
• 13:39
  created
• 13:39
  pilots
• 13:39
  who
• 13:39
  can't
• 13:40
  hand-fly
• 13:40
  in
• 13:40
  emergencies.
• 13:42
  "The
• 13:42
  dangerous
• 13:42
  part
• 13:43
  is
• 13:43
  that
• 13:43
  underload
• 13:44
  feels
• 13:44
  great.
• 13:45
  You're
• 13:45
  productive.
• 13:46
  Things
• 13:46
  are
• 13:46
  shipping.
• 13:47
  The
• 13:47
  feedback
• 13:48
  loop
• 13:48
  is
• 13:48
  all
• 13:48
  positive
• 13:49
  —
• 13:50
  right
• 13:50
  up
• 13:50
  until
• 13:50
  the
• 13:50
  moment
• 13:51
  you
• 13:51
  need
• 13:51
  to
• 13:51
  debug
• 13:51
  something
• 13:52
  and
• 13:52
  realize
• 13:53
  you
• 13:53
  have
• 13:53
  no
• 13:53
  intuition
• 13:54
  about
• 13:54
  how
• 13:54
  the
• 13:54
  system
• 13:55
  behaves."
• 13:56
  Vector
• 13:57
  databases
• 13:57
  continue
• 13:58
  finding
• 13:58
  novel
• 13:58
  applications
• 13:59
  beyond
• 14:00
  chatbot
• 14:00
  RAG
• 14:00
  —
• 14:01
  RuVector's
• 14:01
  DNA
• 14:02
  sequence
• 14:02
  analysis
• 14:03
  example
• 14:03
  demonstrates
• 14:04
  embedding-based
• 14:05
  similarity
• 14:05
  search
• 14:06
  applied
• 14:06
  to
• 14:06
  genomic
• 14:07
  subsequences,
• 14:08
  where
• 14:08
  structural
• 14:08
  similarity
• 14:09
  queries
• 14:09
  replace
• 14:10
  exact
• 14:10
  matching
• 14:10
  for
• 14:10
  bioinformatics
• 14:11
  applications.
• 14:15
  Shanghai
• 14:15
  AI
• 14:16
  Laboratory's
• 14:16
  TrinityGuard
• 14:17
  addresses
• 14:17
  the
• 14:18
  agent
• 14:18
  safety
• 14:18
  gap
• 14:19
  with
• 14:19
  a
• 14:19
  framework
• 14:19
  covering
• 14:20
  20
• 14:20
  risk
• 14:20
  types
• 14:20
  across
• 14:21
  three
• 14:21
  levels:
• 14:21
  single-agent
• 14:22
  risks
• 14:22
  (jailbreak,
• 14:23
  prompt
• 14:23
  injection,
• 14:24
  hallucination,
• 14:25
  memory
• 14:25
  poisoning,
• 14:26
  tool
• 14:26
  misuse),
• 14:27
  inter-agent
• 14:27
  communication
• 14:28
  risks
• 14:28
  (message
• 14:29
  tampering,
• 14:29
  malicious
• 14:30
  propagation,
• 14:30
  goal
• 14:31
  drift,
• 14:31
  identity
• 14:31
  spoofing),
• 14:32
  and
• 14:32
  system-level
• 14:33
  risks
• 14:33
  (cascading
• 14:34
  failures,
• 14:34
  sandbox
• 14:35
  escape,
• 14:35
  rogue
• 14:36
  agents).
• 14:37
  Both
• 14:37
  pre-deployment
• 14:37
  testing
• 14:38
  and
• 14:38
  progressive
• 14:38
  runtime
• 14:39
  monitoring
• 14:39
  are
• 14:39
  supported,
• 14:40
  with
• 14:40
  an
• 14:40
  LLM-powered
• 14:41
  global
• 14:41
  coordinator
• 14:42
  that
• 14:42
  dynamically
• 14:42
  activates
• 14:43
  specific
• 14:43
  risk
• 14:44
  monitors
• 14:44
  based
• 14:44
  on
• 14:44
  conversation
• 14:45
  context.
• 14:46
  Built
• 14:46
  on
• 14:46
  AG2/AutoGen
• 14:48
  and
• 14:48
  aligned
• 14:49
  with
• 14:49
  the
• 14:49
  OWASP
• 14:49
  Agentic
• 14:50
  AI
• 14:50
  Security
• 14:50
  Top
• 14:50
  10,
• 14:51
  it
• 14:51
  provides
• 14:51
  the
• 14:51
  systematic
• 14:52
  safety
• 14:52
  testing
• 14:53
  that
• 14:53
  enterprise
• 14:53
  deployments
• 14:54
  need
• 14:54
  before
• 14:54
  agents
• 14:54
  touch
• 14:55
  production
• 14:55
  data.
• 14:56
  HackMyClaw
• 14:56
  demonstrates
• 14:57
  why
• 14:57
  such
• 14:58
  frameworks
• 14:58
  matter.
• 14:59
  The
• 14:59
  adversarial
• 14:59
  challenge
• 15:00
  pits
• 15:00
  participants
• 15:01
  against
• 15:01
  "Fiu,"
• 15:02
  a
• 15:02
  Claude
• 15:02
  Opus
• 15:02
  4.6-powered
• 15:03
  assistant
• 15:04
  with
• 15:04
  prompt-only
• 15:05
  security
• 15:05
  instructions
• 15:05
  (10–20
• 15:06
  lines
• 15:07
  telling
• 15:07
  it
• 15:07
  never
• 15:07
  to
• 15:08
  reveal
• 15:08
  `secrets.env`).
• 15:10
  Participants
• 15:10
  send
• 15:11
  emails
• 15:11
  with
• 15:11
  prompt
• 15:11
  injection
• 15:12
  payloads
• 15:12
  —
• 15:13
  invisible
• 15:13
  Unicode
• 15:14
  characters,
• 15:14
  multi-step
• 15:15
  reasoning
• 15:15
  exploits,
• 15:16
  social
• 15:16
  engineering
• 15:17
  —
• 15:17
  and
• 15:17
  successful
• 15:18
  extraction
• 15:18
  of
• 15:19
  API
• 15:19
  keys
• 15:19
  earns
• 15:20
  $300.
• 15:21
  The
• 15:21
  creator's
• 15:22
  expectation
• 15:22
  is
• 15:23
  explicit:
• 15:23
  state-of-the-art
• 15:24
  models
• 15:24
  are
• 15:25
  not
• 15:25
  unhackable.
• 15:26
  Google
• 15:26
  has
• 15:26
  already
• 15:27
  had
• 15:27
  to
• 15:27
  reinstate
• 15:27
  the
• 15:27
  bot's
• 15:28
  email
• 15:28
  after
• 15:28
  takedown.
• 15:29
  A
• 15:29
  viral
• 15:30
  video
• 15:30
  walkthrough
• 15:31
  of
• 15:31
  OpenClaw
• 15:31
  shows
• 15:32
  the
• 15:32
  other
• 15:32
  end
• 15:32
  of
• 15:32
  the
• 15:32
  spectrum:
• 15:33
  giving
• 15:33
  an
• 15:33
  LLM
• 15:34
  full
• 15:34
  control
• 15:35
  of
• 15:35
  a
• 15:35
  Kali
• 15:35
  Linux
• 15:35
  machine,
• 15:36
  accessible
• 15:37
  via
• 15:37
  Telegram.
• 15:38
  The
• 15:38
  demonstration
• 15:39
  configures
• 15:39
  Claude
• 15:39
  Opus
• 15:40
  4.6
• 15:41
  on
• 15:41
  a
• 15:41
  cloud-hosted
• 15:42
  Kali
• 15:42
  instance,
• 15:43
  then
• 15:43
  tasks
• 15:44
  it
• 15:44
  entirely
• 15:44
  by
• 15:44
  phone:
• 15:45
  finding
• 15:46
  nearby
• 15:46
  CCTV
• 15:47
  cameras,
• 15:48
  running
• 15:48
  OSINT
• 15:48
  on
• 15:49
  named
• 15:49
  individuals,
• 15:50
  and
• 15:50
  performing
• 15:51
  automated
• 15:51
  pentests
• 15:52
  using
• 15:52
  sub-agents
• 15:53
  that
• 15:53
  independently
• 15:54
  install
• 15:54
  tools,
• 15:55
  select
• 15:55
  models,
• 15:56
  and
• 15:56
  produce
• 15:56
  vulnerability
• 15:57
  reports
• 15:57
  —
• 15:58
  including
• 15:59
  cracked
• 15:59
  passwords
• 16:00
  and
• 16:00
  SQL
• 16:00
  injection
• 16:01
  payloads
• 16:01
  —
• 16:02
  without
• 16:02
  the
• 16:02
  user
• 16:02
  ever
• 16:03
  opening
• 16:03
  a
• 16:03
  laptop.
• 16:04
  The
• 16:04
  framework's
• 16:05
  community
• 16:05
  skill
• 16:06
  hub
• 16:06
  already
• 16:07
  flags
• 16:07
  12%
• 16:08
  of
• 16:08
  submitted
• 16:08
  skills
• 16:09
  as
• 16:09
  malicious,
• 16:10
  underscoring
• 16:11
  that
• 16:11
  the
• 16:11
  surface
• 16:11
  area
• 16:12
  for
• 16:12
  agentic
• 16:12
  exploitation
• 16:13
  is
• 16:13
  expanding
• 16:14
  faster
• 16:14
  than
• 16:14
  the
• 16:14
  guardrails.
• 16:17
  A
• 16:17
  forensic
• 16:18
  teardown
• 16:18
  of
• 16:18
  LinkedIn's
• 16:19
  identity
• 16:19
  verification
• 16:20
  reveals
• 16:20
  the
• 16:21
  true
• 16:21
  cost
• 16:21
  of
• 16:21
  a
• 16:22
  blue
• 16:22
  checkmark.
• 16:23
  When
• 16:23
  users
• 16:23
  tap
• 16:23
  "verify,"
• 16:24
  their
• 16:24
  passport,
• 16:25
  selfie,
• 16:25
  and
• 16:26
  biometric
• 16:26
  data
• 16:26
  go
• 16:27
  not
• 16:27
  to
• 16:27
  LinkedIn
• 16:27
  but
• 16:27
  to
• 16:28
  Persona
• 16:28
  Identities,
• 16:28
  Inc.
• 16:29
  —
• 16:29
  a
• 16:29
  San
• 16:29
  Francisco
• 16:30
  intermediary
• 16:30
  most
• 16:31
  users
• 16:31
  have
• 16:31
  never
• 16:31
  heard
• 16:31
  of.
• 16:32
  The
• 16:32
  34
• 16:33
  pages
• 16:33
  of
• 16:33
  legal
• 16:33
  documentation
• 16:34
  disclose
• 16:34
  that
• 16:35
  Persona
• 16:35
  collects
• 16:35
  full
• 16:35
  passport
• 16:36
  data
• 16:36
  including
• 16:37
  NFC
• 16:37
  chip
• 16:37
  contents,
• 16:38
  facial
• 16:38
  geometry
• 16:39
  from
• 16:39
  selfie
• 16:39
  and
• 16:39
  passport
• 16:40
  photo,
• 16:41
  behavioral
• 16:41
  biometrics
• 16:41
  (keystroke
• 16:42
  timing,
• 16:42
  interaction
• 16:43
  pauses),
• 16:43
  and
• 16:43
  cross-references
• 16:44
  users
• 16:44
  against
• 16:45
  credit
• 16:45
  agencies
• 16:46
  and
• 16:46
  government
• 16:46
  databases.
• 16:47
  All
• 16:48
  16
• 16:48
  of
• 16:48
  Persona's
• 16:48
  subprocessors
• 16:49
  are
• 16:49
  US-based
• 16:50
  —
• 16:50
  not
• 16:51
  a
• 16:51
  single
• 16:51
  EU
• 16:51
  entity
• 16:52
  in
• 16:52
  the
• 16:52
  chain.
• 16:53
  Under
• 16:53
  the
• 16:53
  CLOUD
• 16:53
  Act,
• 16:54
  US
• 16:54
  law
• 16:54
  enforcement
• 16:55
  can
• 16:55
  compel
• 16:55
  disclosure
• 16:56
  regardless
• 16:56
  of
• 16:56
  where
• 16:56
  data
• 16:57
  physically
• 16:57
  sits;
• 16:58
  Persona's
• 16:58
  own
• 16:59
  privacy
• 16:59
  policy
• 16:59
  confirms
• 17:00
  compliance
• 17:00
  with
• 17:01
  "law
• 17:01
  enforcement,
• 17:01
  national
• 17:02
  security,
• 17:02
  or
• 17:03
  other
• 17:03
  government
• 17:03
  agencies."
• 17:04
  Liability
• 17:05
  for
• 17:05
  a
• 17:05
  breach
• 17:05
  is
• 17:05
  capped
• 17:05
  at
• 17:06
  $50.
• 17:07
  The
• 17:07
  author's
• 17:07
  verdict:
• 17:08
  "I
• 17:08
  handed
• 17:08
  a
• 17:08
  US
• 17:09
  company
• 17:09
  my
• 17:09
  passport,
• 17:10
  my
• 17:10
  face,
• 17:11
  and
• 17:11
  the
• 17:11
  mathematical
• 17:11
  geometry
• 17:12
  of
• 17:12
  my
• 17:12
  skull.
• 17:13
  All
• 17:13
  for
• 17:13
  a
• 17:13
  small
• 17:14
  blue
• 17:14
  checkmark
• 17:14
  on
• 17:14
  a
• 17:14
  professional
• 17:15
  networking
• 17:15
  site."
• 17:16
  The
• 17:16
  erosion
• 17:17
  extends
• 17:17
  across
• 17:17
  platforms.
• 17:18
  Reports
• 17:19
  surfaced
• 17:19
  that
• 17:19
  Grok
• 17:20
  users'
• 17:20
  "deleted"
• 17:20
  conversation
• 17:21
  data
• 17:21
  may
• 17:21
  still
• 17:22
  be
• 17:22
  publicly
• 17:22
  accessible,
• 17:23
  raising
• 17:24
  persistent
• 17:24
  questions
• 17:25
  about
• 17:25
  what
• 17:25
  deletion
• 17:25
  actually
• 17:26
  means
• 17:26
  on
• 17:26
  AI-powered
• 17:27
  platforms.
• 17:28
  Separately,
• 17:29
  users
• 17:29
  investigating
• 17:30
  Gemini's
• 17:30
  behavior
• 17:31
  report
• 17:31
  that
• 17:31
  it
• 17:31
  appears
• 17:32
  to
• 17:32
  recall
• 17:32
  details
• 17:32
  from
• 17:33
  prior
• 17:33
  conversations
• 17:34
  it
• 17:34
  should
• 17:34
  have
• 17:34
  no
• 17:34
  access
• 17:35
  to
• 17:35
  —
• 17:36
  prompting
• 17:36
  speculation
• 17:37
  that
• 17:37
  Google
• 17:37
  may
• 17:37
  be
• 17:37
  running
• 17:38
  an
• 17:38
  undisclosed
• 17:38
  RAG
• 17:38
  layer
• 17:39
  enabling
• 17:39
  cross-session
• 17:40
  memory,
• 17:40
  a
• 17:40
  capability
• 17:41
  that
• 17:41
  blurs
• 17:42
  the
• 17:42
  already
• 17:42
  thin
• 17:42
  boundary
• 17:43
  between
• 17:43
  forgetting
• 17:43
  and
• 17:43
  never-deleting.