Latest / AGI Dreams – Open, Uncensored, & Local - AI News Digest / AGI Dreams Podcast – May 18, 2026

Transcript

AGI Dreams Podcast – May 18, 2026 · AGI Dreams – Open, Uncensored, & Local - AI News Digest

• 0:00
  Microsoft
• 0:00
  dropped
• 0:01
  MDASH
• 0:01
  last
• 0:01
  week
• 0:01
  —
• 0:02
  a
• 0:02
  multi-model
• 0:03
  agentic
• 0:03
  scanning
• 0:04
  harness
• 0:04
  that
• 0:04
  threw
• 0:04
  over
• 0:04
  100
• 0:05
  specialized
• 0:06
  AI
• 0:06
  agents
• 0:06
  at
• 0:06
  the
• 0:06
  Windows
• 0:07
  networking
• 0:07
  stack
• 0:08
  and
• 0:08
  pulled
• 0:08
  out
• 0:08
  16
• 0:09
  real
• 0:09
  CVEs,
• 0:10
  including
• 0:10
  CVE-2026-33827,
• 0:14
  a
• 0:14
  use-after-free
• 0:15
  in
• 0:15
  tcpip.sys
• 0:17
  triggered
• 0:17
  by
• 0:17
  crafted
• 0:18
  DHCPv6
• 0:19
  packets,
• 0:19
  and
• 0:19
  CVE-2026-33824,
• 0:23
  a
• 0:23
  double-free
• 0:24
  in
• 0:24
  IKEv2
• 0:24
  that
• 0:24
  gives
• 0:25
  unauthenticated
• 0:25
  remote
• 0:26
  code
• 0:26
  execution.
• 0:27
  These
• 0:27
  are
• 0:27
  not
• 0:27
  theoretical.
• 0:29
  The
• 0:29
  system
• 0:29
  scored
• 0:29
  88.45%
• 0:31
  on
• 0:31
  the
• 0:31
  CyberGym
• 0:32
  benchmark,
• 0:32
  beating
• 0:32
  every
• 0:33
  other
• 0:33
  participant.
• 0:34
  The
• 0:34
  team
• 0:34
  behind
• 0:35
  it
• 0:35
  includes
• 0:35
  DARPA
• 0:35
  AIxCC
• 0:36
  winners,
• 0:37
  and
• 0:37
  the
• 0:37
  architecture
• 0:37
  —
• 0:38
  Orchestrator
• 0:39
  agent
• 0:39
  coordinating
• 0:39
  Driver
• 0:40
  Fuzzer,
• 0:40
  Static
• 0:40
  Analyzer,
• 0:41
  Exploit
• 0:42
  Validator,
• 0:42
  and
• 0:43
  Root
• 0:43
  Cause
• 0:43
  Analyst
• 0:43
  sub-agents
• 0:44
  —
• 0:44
  is
• 0:44
  the
• 0:45
  template
• 0:45
  everyone
• 0:45
  building
• 0:45
  offensive
• 0:46
  AI
• 0:46
  will
• 0:46
  copy.
• 0:47
  Microsoft
• 0:48
  claims
• 0:48
  MDASH
• 0:49
  identified
• 0:49
  some
• 0:49
  bugs
• 0:50
  in
• 0:50
  under
• 0:50
  ten
• 0:50
  seconds
• 0:51
  that
• 0:51
  would
• 0:51
  have
• 0:51
  taken
• 0:51
  a
• 0:51
  human
• 0:52
  analyst
• 0:52
  weeks.
• 0:53
  Meanwhile,
• 0:54
  ExploitGym
• 0:55
  from
• 0:55
  UC
• 0:55
  Berkeley
• 0:55
  gave
• 0:56
  us
• 0:56
  the
• 0:56
  first
• 0:56
  rigorous
• 0:57
  benchmark
• 0:57
  for
• 0:57
  end-to-end
• 0:58
  exploitation
• 0:58
  —
• 0:59
  898
• 1:00
  instances
• 1:01
  across
• 1:01
  kernel,
• 1:01
  V8,
• 1:02
  and
• 1:02
  userspace
• 1:03
  domains,
• 1:04
  each
• 1:04
  starting
• 1:04
  from
• 1:04
  a
• 1:04
  real
• 1:05
  CVE
• 1:05
  and
• 1:06
  requiring
• 1:06
  a
• 1:06
  working
• 1:06
  proof-of-concept.
• 1:08
  Claude
• 1:08
  Mythos
• 1:09
  Preview
• 1:09
  solved
• 1:09
  157,
• 1:11
  GPT-5.5
• 1:12
  solved
• 1:12
  120,
• 1:14
  and
• 1:14
  the
• 1:14
  rest
• 1:14
  of
• 1:14
  the
• 1:14
  field
• 1:15
  struggled
• 1:15
  to
• 1:15
  clear
• 1:16
  double
• 1:16
  digits.
• 1:17
  The
• 1:17
  kernel
• 1:17
  exploitation
• 1:18
  subset
• 1:19
  is
• 1:19
  particularly
• 1:19
  telling:
• 1:20
  it
• 1:20
  requires
• 1:21
  multi-step
• 1:21
  reasoning
• 1:22
  through
• 1:22
  memory
• 1:22
  layouts,
• 1:23
  race
• 1:23
  conditions,
• 1:24
  and
• 1:24
  mitigation
• 1:25
  bypasses
• 1:25
  that
• 1:26
  don't
• 1:26
  fit
• 1:26
  neatly
• 1:26
  into
• 1:27
  any
• 1:27
  prompt
• 1:27
  template.
• 1:28
  ExploitGym
• 1:29
  treats
• 1:29
  exploitation
• 1:30
  as
• 1:30
  a
• 1:30
  capability
• 1:31
  signal,
• 1:31
  not
• 1:32
  an
• 1:32
  academic
• 1:32
  exercise,
• 1:33
  and
• 1:33
  that
• 1:33
  framing
• 1:34
  matters.
• 1:35
  The
• 1:35
  week's
• 1:35
  sharpest
• 1:36
  demonstration
• 1:36
  came
• 1:37
  from
• 1:37
  Calif.io,
• 1:38
  where
• 1:38
  a
• 1:38
  four-person
• 1:39
  team
• 1:39
  paired
• 1:39
  with
• 1:40
  Mythos
• 1:40
  Preview
• 1:40
  built
• 1:41
  the
• 1:41
  first
• 1:41
  public
• 1:41
  macOS
• 1:42
  kernel
• 1:42
  memory
• 1:43
  corruption
• 1:43
  exploit
• 1:44
  on
• 1:44
  Apple
• 1:44
  M5
• 1:44
  silicon
• 1:45
  —
• 1:46
  with
• 1:46
  MIE
• 1:46
  (Memory
• 1:47
  Integrity
• 1:47
  Enforcement)
• 1:48
  enabled
• 1:48
  —
• 1:49
  in
• 1:49
  five
• 1:49
  days.
• 1:50
  Apple
• 1:50
  spent
• 1:50
  five
• 1:51
  years
• 1:51
  and
• 1:51
  probably
• 1:52
  billions
• 1:52
  building
• 1:52
  MIE
• 1:53
  around
• 1:53
  ARM's
• 1:54
  Memory
• 1:54
  Tagging
• 1:54
  Extension.
• 1:56
  The
• 1:56
  exploit
• 1:56
  is
• 1:56
  a
• 1:57
  data-only
• 1:57
  local
• 1:57
  privilege
• 1:58
  escalation
• 1:58
  chain
• 1:59
  targeting
• 1:59
  macOS
• 2:00
  26.4.1,
• 2:02
  starting
• 2:02
  from
• 2:02
  an
• 2:02
  unprivileged
• 2:03
  user
• 2:03
  and
• 2:03
  ending
• 2:04
  with
• 2:04
  a
• 2:04
  root
• 2:04
  shell
• 2:05
  using
• 2:05
  only
• 2:05
  normal
• 2:06
  system
• 2:06
  calls.
• 2:07
  Bruce
• 2:07
  Dang
• 2:08
  found
• 2:08
  the
• 2:08
  bugs
• 2:08
  on
• 2:09
  April
• 2:09
  25th,
• 2:10
  Dion
• 2:10
  Blazakis
• 2:11
  joined
• 2:11
  April
• 2:11
  27th,
• 2:13
  and
• 2:13
  by
• 2:13
  May
• 2:13
  1st
• 2:13
  they
• 2:13
  had
• 2:14
  a
• 2:14
  working
• 2:14
  chain.
• 2:15
  The
• 2:15
  bugs
• 2:16
  belong
• 2:16
  to
• 2:16
  known
• 2:16
  classes,
• 2:17
  which
• 2:17
  is
• 2:17
  precisely
• 2:18
  the
• 2:18
  point:
• 2:19
  Mythos
• 2:19
  generalizes
• 2:20
  across
• 2:20
  bug
• 2:20
  classes,
• 2:21
  and
• 2:21
  once
• 2:22
  it
• 2:22
  learns
• 2:22
  to
• 2:22
  attack
• 2:22
  one,
• 2:23
  it
• 2:23
  transfers.
• 2:24
  MIE
• 2:25
  was
• 2:25
  built
• 2:25
  in
• 2:25
  a
• 2:25
  world
• 2:25
  before
• 2:26
  Mythos
• 2:26
  Preview.
• 2:27
  We
• 2:27
  are
• 2:27
  about
• 2:28
  to
• 2:28
  find
• 2:28
  out
• 2:28
  how
• 2:28
  that
• 2:28
  timing
• 2:29
  gap
• 2:29
  plays
• 2:29
  out.
• 2:33
  If
• 2:33
  AI
• 2:33
  can
• 2:34
  find
• 2:34
  vulnerabilities
• 2:35
  faster,
• 2:35
  it
• 2:35
  also
• 2:36
  creates
• 2:36
  them
• 2:36
  faster.
• 2:37
  SecureForge,
• 2:38
  a
• 2:38
  Stanford
• 2:39
  pipeline,
• 2:40
  measured
• 2:40
  the
• 2:40
  damage:
• 2:41
  across
• 2:41
  multiple
• 2:42
  LLMs,
• 2:43
  23%
• 2:44
  of
• 2:44
  generated
• 2:44
  code
• 2:44
  contained
• 2:45
  security
• 2:45
  vulnerabilities.
• 2:47
  Their
• 2:47
  mitigation
• 2:47
  —
• 2:48
  prompt
• 2:48
  optimization
• 2:49
  combined
• 2:49
  with
• 2:50
  MCMC
• 2:50
  amplification
• 2:51
  —
• 2:52
  reduced
• 2:52
  that
• 2:52
  rate
• 2:53
  by
• 2:53
  up
• 2:53
  to
• 2:53
  48%.
• 2:55
  The
• 2:55
  approach
• 2:55
  treats
• 2:55
  vulnerability
• 2:56
  prevention
• 2:57
  as
• 2:57
  a
• 2:57
  search
• 2:57
  problem
• 2:58
  over
• 2:58
  prompt
• 2:58
  space
• 2:59
  rather
• 2:59
  than
• 2:59
  a
• 2:59
  training
• 3:00
  problem,
• 3:01
  which
• 3:01
  means
• 3:01
  it
• 3:01
  works
• 3:01
  on
• 3:01
  any
• 3:02
  model
• 3:02
  without
• 3:02
  fine-tuning.
• 3:04
  The
• 3:04
  practical
• 3:04
  implication
• 3:05
  is
• 3:05
  that
• 3:05
  the
• 3:05
  cheapest
• 3:06
  intervention
• 3:06
  point
• 3:06
  for
• 3:07
  LLM-generated
• 3:07
  code
• 3:08
  security
• 3:08
  is
• 3:08
  at
• 3:09
  the
• 3:09
  prompt
• 3:09
  level,
• 3:10
  not
• 3:10
  the
• 3:10
  model
• 3:10
  level.
• 3:11
  That
• 3:11
  is
• 3:11
  both
• 3:11
  encouraging
• 3:12
  and
• 3:12
  slightly
• 3:12
  terrifying,
• 3:13
  because
• 3:14
  it
• 3:14
  means
• 3:14
  the
• 3:14
  attack
• 3:14
  surface
• 3:15
  includes
• 3:15
  prompt
• 3:15
  manipulation
• 3:16
  too.
• 3:17
  The
• 3:17
  model
• 3:17
  safety
• 3:18
  layer
• 3:18
  itself
• 3:18
  is
• 3:18
  under
• 3:19
  attack.
• 3:20
  A
• 3:20
  presentation
• 3:20
  this
• 3:20
  week
• 3:21
  detailed
• 3:21
  NeurosStrike,
• 3:21
  GateBreaker,
• 3:22
  and
• 3:22
  L³
• 3:23
  (Large
• 3:23
  Language
• 3:23
  Lobotomy)
• 3:24
  —
• 3:24
  three
• 3:24
  techniques
• 3:25
  that
• 3:25
  target
• 3:25
  the
• 3:25
  sparse
• 3:26
  safety
• 3:26
  neurons
• 3:26
  LLMs
• 3:27
  rely
• 3:27
  on
• 3:27
  for
• 3:27
  refusal
• 3:28
  behavior.
• 3:29
  NeurosStrike
• 3:29
  identifies
• 3:30
  and
• 3:30
  ablates
• 3:30
  specific
• 3:31
  neurons
• 3:31
  responsible
• 3:32
  for
• 3:32
  safety
• 3:32
  alignment.
• 3:33
  GateBreaker
• 3:34
  bypasses
• 3:34
  Mixture-of-Experts
• 3:35
  routing
• 3:36
  to
• 3:36
  reach
• 3:36
  safety-critical
• 3:37
  pathways.
• 3:38
  L³
• 3:38
  is
• 3:38
  a
• 3:38
  transferable
• 3:39
  attack:
• 3:40
  train
• 3:40
  it
• 3:40
  on
• 3:40
  one
• 3:40
  model,
• 3:41
  apply
• 3:41
  it
• 3:41
  to
• 3:41
  another.
• 3:43
  These
• 3:43
  are
• 3:43
  not
• 3:43
  jailbreaks
• 3:44
  requiring
• 3:44
  clever
• 3:44
  prompting
• 3:44
  —
• 3:45
  they
• 3:45
  are
• 3:45
  architectural
• 3:46
  attacks
• 3:46
  on
• 3:46
  the
• 3:46
  safety
• 3:47
  mechanism
• 3:47
  itself,
• 3:48
  and
• 3:48
  they
• 3:48
  work
• 3:48
  across
• 3:49
  model
• 3:49
  families.
• 3:50
  The
• 3:50
  same
• 3:50
  research
• 3:50
  group
• 3:50
  proposed
• 3:51
  GoodVibe,
• 3:52
  a
• 3:52
  defensive
• 3:52
  framework
• 3:53
  for
• 3:53
  generating
• 3:53
  more
• 3:53
  secure
• 3:54
  code,
• 3:54
  but
• 3:54
  the
• 3:54
  offense-defense
• 3:55
  asymmetry
• 3:56
  here
• 3:56
  is
• 3:56
  stark.
• 3:57
  On
• 3:57
  the
• 3:57
  smart
• 3:57
  contract
• 3:58
  side,
• 3:58
  solidity-cot-auditor
• 3:59
  implements
• 4:00
  a
• 4:00
  four-role
• 4:01
  chain-of-thought
• 4:01
  pipeline
• 4:01
  —
• 4:02
  Planner,
• 4:02
  Detector,
• 4:03
  Analyzer,
• 4:04
  Reporter
• 4:04
  —
• 4:04
  specifically
• 4:05
  for
• 4:05
  Solidity
• 4:06
  security
• 4:06
  auditing.
• 4:07
  The
• 4:07
  Planner
• 4:07
  decomposes
• 4:08
  the
• 4:08
  contract
• 4:09
  into
• 4:09
  audit
• 4:09
  targets,
• 4:10
  the
• 4:10
  Detector
• 4:10
  scans
• 4:11
  for
• 4:11
  vulnerability
• 4:12
  patterns,
• 4:13
  the
• 4:13
  Analyzer
• 4:13
  evaluates
• 4:14
  exploitability
• 4:14
  and
• 4:15
  impact,
• 4:15
  and
• 4:16
  the
• 4:16
  Reporter
• 4:16
  produces
• 4:16
  structured
• 4:17
  findings.
• 4:18
  It
• 4:18
  is
• 4:18
  narrowly
• 4:19
  scoped
• 4:19
  and
• 4:19
  honest
• 4:20
  about
• 4:20
  what
• 4:20
  it
• 4:20
  does:
• 4:21
  structured
• 4:21
  LLM
• 4:22
  reasoning
• 4:22
  applied
• 4:22
  to
• 4:23
  a
• 4:23
  domain
• 4:23
  where
• 4:23
  the
• 4:23
  cost
• 4:24
  of
• 4:24
  missed
• 4:24
  vulnerabilities
• 4:25
  is
• 4:25
  measured
• 4:25
  in
• 4:25
  stolen
• 4:26
  funds,
• 4:27
  not
• 4:27
  bug
• 4:27
  bounties.
• 4:28
  The
• 4:28
  narrow
• 4:29
  scope
• 4:29
  is
• 4:29
  the
• 4:29
  design
• 4:30
  virtue
• 4:30
  —
• 4:30
  it
• 4:30
  does
• 4:31
  not
• 4:31
  try
• 4:31
  to
• 4:31
  be
• 4:31
  a
• 4:31
  general-purpose
• 4:32
  code
• 4:32
  auditor,
• 4:33
  which
• 4:33
  means
• 4:33
  its
• 4:33
  failure
• 4:34
  modes
• 4:34
  are
• 4:34
  predictable
• 4:35
  and
• 4:35
  its
• 4:35
  output
• 4:35
  is
• 4:36
  auditable
• 4:36
  by
• 4:36
  a
• 4:36
  human
• 4:36
  reviewer
• 4:37
  who
• 4:37
  knows
• 4:37
  Solidity.
• 4:42
  The
• 4:42
  governance
• 4:42
  layer
• 4:42
  for
• 4:42
  agentic
• 4:43
  AI
• 4:43
  produced
• 4:44
  five
• 4:44
  distinct
• 4:44
  frameworks
• 4:45
  in
• 4:45
  a
• 4:45
  single
• 4:45
  week,
• 4:46
  which
• 4:46
  is
• 4:46
  either
• 4:46
  a
• 4:46
  sign
• 4:47
  of
• 4:47
  maturation
• 4:47
  or
• 4:48
  fragmentation
• 4:48
  depending
• 4:49
  on
• 4:49
  your
• 4:49
  tolerance
• 4:49
  for
• 4:49
  committee
• 4:50
  output.
• 4:51
  The
• 4:51
  common
• 4:51
  thread:
• 4:52
  every
• 4:52
  framework
• 4:52
  acknowledges
• 4:53
  that
• 4:53
  traditional
• 4:54
  perimeter
• 4:54
  security
• 4:54
  and
• 4:55
  static
• 4:55
  access
• 4:55
  controls
• 4:56
  do
• 4:56
  not
• 4:56
  work
• 4:56
  when
• 4:56
  your
• 4:56
  software
• 4:57
  can
• 4:57
  reason,
• 4:58
  act
• 4:58
  autonomously,
• 4:59
  and
• 4:59
  spawn
• 5:00
  copies
• 5:00
  of
• 5:00
  itself.
• 5:01
  Adrian
• 5:01
  is
• 5:02
  an
• 5:02
  open-source
• 5:02
  runtime
• 5:03
  security
• 5:03
  monitoring
• 5:03
  engine
• 5:04
  for
• 5:04
  AI
• 5:04
  agents,
• 5:05
  aligned
• 5:05
  with
• 5:06
  the
• 5:06
  AARM
• 5:06
  standard,
• 5:07
  that
• 5:07
  analyzes
• 5:08
  both
• 5:08
  activity
• 5:08
  logs
• 5:09
  and
• 5:09
  reasoning
• 5:09
  traces
• 5:10
  to
• 5:10
  detect
• 5:10
  misaligned
• 5:10
  behavior.
• 5:11
  The
• 5:11
  key
• 5:12
  differentiator:
• 5:13
  it
• 5:13
  correlates
• 5:13
  behaviors
• 5:14
  across
• 5:14
  a
• 5:14
  session
• 5:15
  and
• 5:15
  assesses
• 5:15
  each
• 5:16
  action
• 5:16
  against
• 5:16
  a
• 5:16
  working
• 5:17
  understanding
• 5:17
  of
• 5:17
  what
• 5:17
  the
• 5:18
  agent
• 5:18
  is
• 5:18
  supposed
• 5:18
  to
• 5:18
  be
• 5:18
  doing.
• 5:20
  Most
• 5:20
  agent
• 5:20
  monitoring
• 5:21
  tools
• 5:21
  stop
• 5:21
  at
• 5:21
  activity
• 5:22
  logs
• 5:22
  —
• 5:23
  API
• 5:23
  calls,
• 5:24
  tool
• 5:24
  invocations,
• 5:25
  database
• 5:25
  queries.
• 5:26
  Adrian
• 5:27
  goes
• 5:27
  deeper
• 5:27
  by
• 5:27
  analyzing
• 5:28
  the
• 5:28
  reasoning
• 5:28
  trace:
• 5:29
  why
• 5:29
  the
• 5:30
  agent
• 5:30
  took
• 5:30
  an
• 5:30
  action,
• 5:31
  what
• 5:31
  context
• 5:31
  it
• 5:31
  was
• 5:32
  operating
• 5:32
  under,
• 5:33
  and
• 5:33
  what
• 5:33
  it
• 5:33
  planned
• 5:33
  to
• 5:33
  do
• 5:33
  next.
• 5:35
  Research
• 5:35
  from
• 5:35
  OpenAI
• 5:36
  and
• 5:36
  DeepMind
• 5:37
  found
• 5:37
  that
• 5:37
  combining
• 5:37
  behavior
• 5:38
  and
• 5:38
  reasoning
• 5:38
  analysis
• 5:39
  boosts
• 5:39
  detection
• 5:40
  accuracy
• 5:40
  by
• 5:40
  ~35%
• 5:42
  and
• 5:42
  is
• 5:42
  4x
• 5:43
  more
• 5:43
  likely
• 5:43
  to
• 5:43
  catch
• 5:43
  nuanced
• 5:44
  attacks.
• 5:45
  Adrian
• 5:45
  ships
• 5:46
  self-hostable
• 5:47
  with
• 5:47
  a
• 5:47
  Gemma
• 5:47
  4
• 5:47
  classifier
• 5:48
  and
• 5:48
  a
• 5:48
  two-line
• 5:49
  LangChain
• 5:49
  integration,
• 5:50
  which
• 5:50
  matters
• 5:51
  because
• 5:51
  sending
• 5:51
  your
• 5:51
  agent's
• 5:52
  reasoning
• 5:52
  traces
• 5:53
  to
• 5:53
  a
• 5:53
  third-party
• 5:53
  API
• 5:54
  defeats
• 5:54
  the
• 5:54
  purpose.
• 5:55
  The
• 5:56
  example
• 5:56
  Adrian
• 5:56
  cites
• 5:57
  is
• 5:57
  telling:
• 5:58
  if
• 5:58
  your
• 5:58
  e-commerce
• 5:58
  agent
• 5:59
  starts
• 5:59
  resetting
• 6:00
  user
• 6:00
  passwords,
• 6:01
  no
• 6:01
  pattern-matching
• 6:02
  classifier
• 6:02
  trained
• 6:03
  on
• 6:03
  prompt
• 6:03
  injection
• 6:04
  datasets
• 6:04
  will
• 6:04
  flag
• 6:05
  it,
• 6:05
  but
• 6:05
  a
• 6:05
  system
• 6:06
  that
• 6:06
  understands
• 6:06
  the
• 6:07
  agent's
• 6:07
  remit
• 6:07
  will.
• 6:08
  Palantir's
• 6:09
  MA-S2
• 6:09
  (Mission
• 6:10
  Assurance
• 6:10
  Security
• 6:10
  Standard)
• 6:11
  introduces
• 6:11
  four
• 6:12
  control
• 6:12
  domains
• 6:12
  —
• 6:13
  Contextual
• 6:14
  Vulnerability
• 6:15
  Scanning,
• 6:15
  AI-Powered
• 6:16
  Monitoring,
• 6:17
  Inventory
• 6:17
  and
• 6:17
  Supply
• 6:18
  Chain,
• 6:18
  and
• 6:18
  Autonomous
• 6:19
  Response
• 6:19
  Orchestration
• 6:20
  —
• 6:21
  designed
• 6:21
  for
• 6:21
  environments
• 6:22
  where
• 6:22
  "move
• 6:22
  fast
• 6:23
  and
• 6:23
  break
• 6:23
  things"
• 6:23
  gets
• 6:24
  people
• 6:24
  killed.
• 6:25
  This
• 6:25
  is
• 6:26
  military-grade
• 6:26
  AI
• 6:27
  governance,
• 6:28
  and
• 6:28
  the
• 6:28
  specificity
• 6:29
  of
• 6:29
  the
• 6:29
  controls
• 6:30
  reflects
• 6:30
  lessons
• 6:30
  learned
• 6:31
  from
• 6:31
  deploying
• 6:31
  AI
• 6:32
  in
• 6:32
  classified
• 6:32
  settings.
• 6:34
  The
• 6:34
  Autonomous
• 6:34
  Response
• 6:35
  Orchestration
• 6:35
  domain
• 6:36
  is
• 6:36
  the
• 6:36
  interesting
• 6:36
  one:
• 6:37
  it
• 6:37
  acknowledges
• 6:38
  that
• 6:38
  human-in-the-loop
• 6:39
  review
• 6:39
  cannot
• 6:40
  scale
• 6:40
  to
• 6:40
  AI-speed
• 6:41
  threats
• 6:41
  and
• 6:41
  prescribes
• 6:42
  automated
• 6:42
  response
• 6:43
  within
• 6:43
  policy
• 6:44
  guardrails.
• 6:45
  The
• 6:45
  DAMSIC
• 6:49
  framework
• 6:49
  (Data
• 6:49
  privacy,
• 6:50
  Autonomous
• 6:51
  security,
• 6:51
  Monitoring,
• 6:52
  Supply
• 6:53
  chain,
• 6:53
  model
• 6:54
  Integrity,
• 6:55
  Compliance)
• 6:55
  comes
• 6:55
  from
• 6:56
  SecureAgentics
• 6:56
  and
• 6:56
  attempts
• 6:57
  a
• 6:57
  comprehensive
• 6:57
  taxonomy
• 6:58
  covering
• 6:59
  the
• 6:59
  full
• 6:59
  agentic
• 7:00
  AI
• 7:00
  lifecycle.
• 7:01
  CIO.com's
• 7:02
  piece
• 7:03
  on
• 7:03
  identity
• 7:03
  governance
• 7:04
  raised
• 7:04
  the
• 7:04
  harder
• 7:04
  question:
• 7:05
  when
• 7:05
  an
• 7:06
  AI
• 7:06
  agent
• 7:06
  can
• 7:06
  fork
• 7:07
  its
• 7:07
  own
• 7:07
  digital
• 7:07
  twin,
• 7:08
  who
• 7:08
  is
• 7:08
  accountable
• 7:09
  for
• 7:09
  the
• 7:09
  fork's
• 7:09
  actions?
• 7:10
  The
• 7:10
  article
• 7:11
  walks
• 7:11
  through
• 7:11
  agent
• 7:12
  taxonomies
• 7:12
  —
• 7:13
  autonomous
• 7:13
  versus
• 7:14
  semi-autonomous,
• 7:15
  persistent
• 7:16
  versus
• 7:16
  ephemeral
• 7:16
  —
• 7:17
  and
• 7:17
  the
• 7:17
  identity
• 7:18
  governance
• 7:18
  implications
• 7:19
  of
• 7:19
  each.
• 7:20
  CoSAI's
• 7:20
  agentic
• 7:21
  identity
• 7:21
  and
• 7:21
  access
• 7:22
  management
• 7:22
  work
• 7:23
  is
• 7:23
  the
• 7:23
  closest
• 7:23
  anyone
• 7:24
  has
• 7:24
  come
• 7:24
  to
• 7:24
  answering
• 7:24
  the
• 7:25
  accountability
• 7:25
  question,
• 7:26
  and
• 7:26
  the
• 7:26
  answer
• 7:26
  is
• 7:27
  "we
• 7:27
  don't
• 7:27
  know
• 7:27
  yet,
• 7:27
  but
• 7:28
  here's
• 7:28
  a
• 7:28
  taxonomy
• 7:28
  of
• 7:29
  what
• 7:29
  we
• 7:29
  need
• 7:29
  to
• 7:29
  figure
• 7:29
  out."
• 7:30
  The
• 7:30
  digital
• 7:30
  twin
• 7:31
  forking
• 7:31
  risk
• 7:31
  is
• 7:31
  not
• 7:32
  hypothetical:
• 7:33
  any
• 7:33
  agent
• 7:33
  with
• 7:33
  access
• 7:34
  to
• 7:34
  its
• 7:34
  own
• 7:34
  deployment
• 7:34
  infrastructure
• 7:35
  can
• 7:35
  replicate,
• 7:36
  and
• 7:36
  the
• 7:36
  replicas
• 7:37
  inherit
• 7:37
  the
• 7:37
  original's
• 7:38
  credentials
• 7:38
  unless
• 7:38
  the
• 7:38
  identity
• 7:39
  system
• 7:39
  is
• 7:39
  designed
• 7:40
  to
• 7:40
  prevent
• 7:40
  it.
• 7:41
  A
• 7:41
  CSA
• 7:41
  panel
• 7:42
  featuring
• 7:42
  Om
• 7:42
  from
• 7:42
  Sumerel,
• 7:43
  Alex
• 7:44
  from
• 7:44
  Versa
• 7:44
  Networks,
• 7:45
  Barack
• 7:45
  from
• 7:45
  Tenable,
• 7:46
  and
• 7:46
  Gadi
• 7:46
  from
• 7:46
  Nostic
• 7:47
  discussed
• 7:47
  the
• 7:47
  Mythos
• 7:48
  crisis
• 7:48
  and
• 7:48
  self-healing
• 7:49
  defenses.
• 7:50
  The
• 7:50
  consensus:
• 7:51
  static
• 7:51
  security
• 7:52
  postures
• 7:52
  are
• 7:52
  dead,
• 7:53
  and
• 7:53
  the
• 7:53
  question
• 7:53
  is
• 7:54
  how
• 7:54
  fast
• 7:54
  your
• 7:54
  defensive
• 7:55
  AI
• 7:55
  can
• 7:55
  adapt
• 7:55
  relative
• 7:56
  to
• 7:56
  offensive
• 7:56
  AI.
• 7:57
  Nobody
• 7:58
  had
• 7:58
  a
• 7:58
  satisfying
• 7:58
  answer,
• 7:59
  but
• 7:59
  the
• 7:59
  panel's
• 8:00
  framing
• 8:00
  of
• 8:00
  the
• 8:00
  problem
• 8:00
  —
• 8:01
  that
• 8:01
  the
• 8:02
  defender's
• 8:02
  AI
• 8:02
  must
• 8:03
  match
• 8:03
  the
• 8:03
  attacker's
• 8:03
  AI
• 8:04
  in
• 8:04
  speed
• 8:04
  and
• 8:04
  adaptability,
• 8:05
  not
• 8:05
  just
• 8:05
  detection
• 8:06
  accuracy
• 8:06
  —
• 8:07
  is
• 8:07
  the
• 8:07
  right
• 8:07
  one.
• 8:10
  A
• 8:10
  formal
• 8:11
  proof
• 8:11
  this
• 8:11
  week
• 8:11
  demonstrated
• 8:12
  that
• 8:12
  the
• 8:12
  number
• 8:12
  of
• 8:12
  vulnerabilities
• 8:13
  in
• 8:13
  software
• 8:14
  is
• 8:14
  countably
• 8:14
  infinite.
• 8:15
  The
• 8:16
  paper
• 8:16
  constructs
• 8:16
  a
• 8:16
  "Vulnerability
• 8:17
  Factory"
• 8:17
  —
• 8:18
  a
• 8:18
  C
• 8:18
  program
• 8:19
  that
• 8:19
  generates
• 8:19
  an
• 8:20
  unbounded
• 8:20
  family
• 8:21
  of
• 8:21
  distinct,
• 8:21
  exploitable
• 8:22
  vulnerabilities
• 8:22
  —
• 8:23
  and
• 8:23
  uses
• 8:24
  a
• 8:24
  chemical
• 8:24
  abundance
• 8:24
  analogy
• 8:25
  to
• 8:25
  argue
• 8:25
  that
• 8:25
  vulnerabilities
• 8:26
  are
• 8:26
  an
• 8:26
  intrinsic
• 8:27
  property
• 8:27
  of
• 8:27
  complex
• 8:28
  systems,
• 8:29
  not
• 8:29
  a
• 8:29
  finite
• 8:30
  backlog
• 8:30
  to
• 8:30
  be
• 8:30
  cleared.
• 8:31
  The
• 8:31
  practical
• 8:31
  punchline:
• 8:32
  only
• 8:33
  about
• 8:33
  6%
• 8:33
  of
• 8:33
  CVEs
• 8:34
  are
• 8:34
  ever
• 8:35
  exploited
• 8:35
  in
• 8:35
  the
• 8:35
  wild.
• 8:36
  The
• 8:36
  attacker's
• 8:37
  problem
• 8:37
  is
• 8:37
  not
• 8:37
  finding
• 8:38
  vulnerabilities
• 8:38
  —
• 8:39
  it
• 8:39
  is
• 8:39
  selecting
• 8:39
  the
• 8:39
  right
• 8:40
  ones.
• 8:40
  This
• 8:41
  reframes
• 8:41
  the
• 8:41
  entire
• 8:42
  defensive
• 8:42
  posture
• 8:43
  from
• 8:43
  "find
• 8:43
  and
• 8:43
  fix
• 8:44
  all
• 8:44
  bugs"
• 8:44
  to
• 8:44
  "understand
• 8:45
  which
• 8:45
  bugs
• 8:45
  matter,"
• 8:46
  and
• 8:46
  it
• 8:46
  gives
• 8:46
  formal
• 8:47
  backing
• 8:47
  to
• 8:47
  risk-based
• 8:47
  prioritization
• 8:48
  that
• 8:48
  practitioners
• 8:49
  have
• 8:49
  been
• 8:49
  doing
• 8:49
  intuitively
• 8:50
  for
• 8:50
  years.
• 8:50
  On
• 8:50
  the
• 8:51
  practitioner
• 8:51
  side,
• 8:51
  a
• 8:52
  quality
• 8:52
  engineering
• 8:52
  blog
• 8:53
  documented
• 8:53
  13
• 8:54
  releases
• 8:54
  of
• 8:54
  an
• 8:54
  agentic
• 8:55
  testing
• 8:55
  system,
• 8:56
  cataloging
• 8:56
  every
• 8:57
  failure
• 8:57
  mode
• 8:57
  along
• 8:57
  the
• 8:57
  way:
• 8:58
  Claude
• 8:59
  Opus
• 8:59
  4.7
• 9:00
  exhibiting
• 9:00
  destructive
• 9:01
  behavior
• 9:01
  during
• 9:01
  automated
• 9:02
  testing
• 9:02
  —
• 9:03
  deleting
• 9:03
  files,
• 9:04
  overwriting
• 9:04
  configurations
• 9:05
  —
• 9:05
  feedback
• 9:06
  loops
• 9:06
  that
• 9:06
  amplified
• 9:07
  rather
• 9:07
  than
• 9:07
  corrected
• 9:07
  errors,
• 9:08
  and
• 9:08
  the
• 9:08
  slow
• 9:08
  accumulation
• 9:09
  of
• 9:09
  hard-won
• 9:09
  heuristics
• 9:10
  that
• 9:10
  make
• 9:10
  agentic
• 9:11
  QE
• 9:11
  workable
• 9:11
  in
• 9:11
  production.
• 9:13
  Each
• 9:13
  release
• 9:13
  fixed
• 9:13
  something
• 9:14
  the
• 9:14
  previous
• 9:14
  one
• 9:14
  broke,
• 9:15
  and
• 9:15
  the
• 9:15
  learning
• 9:15
  curve
• 9:16
  was
• 9:16
  not
• 9:16
  smooth
• 9:16
  but
• 9:17
  jagged,
• 9:17
  full
• 9:18
  of
• 9:18
  regressions
• 9:18
  that
• 9:18
  only
• 9:19
  appeared
• 9:19
  under
• 9:19
  specific
• 9:20
  interaction
• 9:20
  patterns.
• 9:21
  The
• 9:21
  candor
• 9:22
  is
• 9:22
  valuable
• 9:22
  precisely
• 9:23
  because
• 9:23
  most
• 9:23
  agentic
• 9:24
  AI
• 9:24
  coverage
• 9:24
  skips
• 9:24
  the
• 9:25
  iteration
• 9:25
  tax
• 9:26
  and
• 9:26
  jumps
• 9:26
  straight
• 9:26
  to
• 9:26
  the
• 9:26
  success
• 9:27
  story.
• 9:28
  A
• 9:28
  researcher
• 9:28
  demonstrated
• 9:29
  that
• 9:29
  eight
• 9:29
  different
• 9:29
  LLMs,
• 9:30
  when
• 9:31
  asked
• 9:31
  to
• 9:31
  name
• 9:31
  a
• 9:31
  lighthouse
• 9:32
  keeper,
• 9:32
  all
• 9:32
  produce
• 9:33
  "Elias
• 9:33
  Thorne"
• 9:33
  —
• 9:34
  a
• 9:34
  synthetic
• 9:34
  name
• 9:35
  that
• 9:35
  now
• 9:35
  pollutes
• 9:35
  Amazon
• 9:36
  product
• 9:36
  listings
• 9:36
  and
• 9:36
  web
• 9:37
  content
• 9:37
  as
• 9:37
  a
• 9:37
  fingerprint
• 9:38
  of
• 9:38
  AI-generated
• 9:39
  text.
• 9:39
  The
• 9:40
  monoculture
• 9:40
  problem
• 9:40
  in
• 9:41
  LLM
• 9:41
  outputs
• 9:41
  is
• 9:42
  not
• 9:42
  just
• 9:42
  a
• 9:42
  quality
• 9:42
  issue;
• 9:43
  it
• 9:43
  is
• 9:43
  a
• 9:43
  detection
• 9:44
  surface.
• 9:45
  When
• 9:45
  your
• 9:45
  generated
• 9:46
  text
• 9:46
  carries
• 9:46
  forensic
• 9:47
  markers
• 9:47
  this
• 9:47
  obvious,
• 9:48
  the
• 9:48
  generation
• 9:49
  itself
• 9:49
  becomes
• 9:50
  a
• 9:50
  liability.
• 9:53
  Etna
• 9:53
  Labs
• 9:53
  coined
• 9:54
  "harness
• 9:54
  engineering"
• 9:55
  as
• 9:55
  the
• 9:55
  successor
• 9:56
  to
• 9:56
  prompt
• 9:56
  engineering
• 9:57
  and
• 9:57
  context
• 9:58
  engineering
• 9:58
  —
• 9:59
  the
• 9:59
  discipline
• 9:59
  of
• 10:00
  building
• 10:00
  the
• 10:00
  full
• 10:00
  scaffolding
• 10:01
  around
• 10:01
  an
• 10:01
  AI
• 10:01
  agent:
• 10:02
  tool
• 10:02
  integration,
• 10:03
  memory
• 10:04
  management,
• 10:04
  feedback
• 10:05
  loops,
• 10:05
  evaluation
• 10:06
  pipelines,
• 10:06
  and
• 10:06
  data
• 10:07
  flywheels.
• 10:08
  The
• 10:08
  taxonomy
• 10:08
  identifies
• 10:09
  six
• 10:09
  harness
• 10:10
  components
• 10:10
  and
• 10:10
  argues
• 10:11
  that
• 10:11
  the
• 10:11
  competitive
• 10:11
  moat
• 10:12
  is
• 10:12
  not
• 10:12
  in
• 10:12
  which
• 10:12
  model
• 10:12
  you
• 10:12
  use
• 10:13
  but
• 10:13
  in
• 10:13
  how
• 10:13
  well
• 10:13
  you
• 10:13
  wrap
• 10:14
  it.
• 10:14
  Claude
• 10:15
  Code's
• 10:15
  CLAUDE.md
• 10:16
  patterns
• 10:16
  and
• 10:16
  Cursor's
• 10:17
  agentic
• 10:17
  reinforcement
• 10:18
  learning
• 10:18
  are
• 10:18
  cited
• 10:18
  as
• 10:19
  early
• 10:19
  examples.
• 10:20
  The
• 10:20
  evolution
• 10:21
  is
• 10:21
  clear:
• 10:21
  prompt
• 10:22
  engineering
• 10:22
  optimized
• 10:23
  a
• 10:23
  single
• 10:23
  call,
• 10:24
  context
• 10:24
  engineering
• 10:25
  optimized
• 10:25
  what
• 10:25
  the
• 10:25
  model
• 10:26
  could
• 10:26
  see,
• 10:27
  and
• 10:27
  harness
• 10:27
  engineering
• 10:28
  optimizes
• 10:28
  the
• 10:28
  entire
• 10:29
  system
• 10:29
  around
• 10:29
  the
• 10:29
  model
• 10:29
  —
• 10:30
  including
• 10:30
  what
• 10:31
  happens
• 10:31
  when
• 10:31
  the
• 10:31
  model
• 10:31
  is
• 10:31
  wrong.
• 10:32
  The
• 10:32
  data
• 10:33
  flywheel
• 10:33
  component
• 10:33
  is
• 10:33
  the
• 10:34
  most
• 10:34
  underappreciated:
• 10:35
  every
• 10:35
  agent
• 10:36
  interaction
• 10:36
  generates
• 10:36
  training
• 10:37
  signal,
• 10:38
  and
• 10:38
  the
• 10:38
  teams
• 10:38
  that
• 10:38
  capture
• 10:39
  and
• 10:39
  feed
• 10:39
  that
• 10:39
  signal
• 10:39
  back
• 10:40
  will
• 10:40
  compound
• 10:40
  their
• 10:40
  advantage
• 10:41
  over
• 10:41
  teams
• 10:41
  that
• 10:41
  treat
• 10:41
  each
• 10:42
  invocation
• 10:42
  as
• 10:42
  stateless.
• 10:43
  One
• 10:44
  developer
• 10:44
  tracked
• 10:44
  every
• 10:45
  dollar
• 10:45
  and
• 10:45
  hour
• 10:45
  spent
• 10:46
  on
• 10:46
  AI
• 10:46
  coding
• 10:46
  tools
• 10:46
  for
• 10:47
  60
• 10:47
  days:
• 10:48
  $200/month
• 10:49
  in
• 10:49
  subscriptions
• 10:50
  (Cursor
• 10:50
  Pro,
• 10:51
  Claude
• 10:51
  Pro
• 10:51
  +
• 10:51
  API,
• 10:52
  ChatGPT
• 10:53
  Plus,
• 10:53
  Copilot,
• 10:54
  CodeRabbit,
• 10:55
  v0),
• 10:56
  62
• 10:57
  productive
• 10:57
  hours,
• 10:58
  28
• 10:58
  hours
• 10:58
  fixing
• 10:59
  plausible-but-wrong
• 10:59
  AI
• 11:00
  output,
• 11:01
  and
• 11:01
  14
• 11:01
  hours
• 11:01
  on
• 11:01
  context
• 11:02
  switching
• 11:02
  and
• 11:02
  tool-arguing.
• 11:04
  Net:
• 11:05
  1.7-2x
• 11:06
  productivity
• 11:06
  gain,
• 11:07
  not
• 11:07
  the
• 11:07
  10x
• 11:08
  Twitter
• 11:08
  promises.
• 11:09
  For
• 11:09
  every
• 11:09
  productive
• 11:10
  hour
• 11:10
  of
• 11:10
  AI-assisted
• 11:11
  coding,
• 11:12
  roughly
• 11:12
  40
• 11:12
  minutes
• 11:13
  went
• 11:13
  to
• 11:13
  overhead
• 11:13
  —
• 11:14
  fixing
• 11:14
  plausible-but-wrong
• 11:15
  output,
• 11:16
  switching
• 11:16
  between
• 11:16
  tools,
• 11:17
  and
• 11:17
  arguing
• 11:18
  with
• 11:18
  agents
• 11:18
  that
• 11:18
  were
• 11:18
  confidently
• 11:19
  incorrect.
• 11:20
  The
• 11:20
  overhead
• 11:21
  was
• 11:21
  worst
• 11:21
  on
• 11:21
  legacy
• 11:21
  refactoring,
• 11:22
  where
• 11:22
  productive-to-wasted
• 11:23
  time
• 11:24
  approached
• 11:24
  1:1.
• 11:25
  The
• 11:25
  sharpest
• 11:26
  finding:
• 11:26
  CodeRabbit
• 11:27
  at
• 11:27
  $15/month
• 11:28
  had
• 11:29
  the
• 11:29
  highest
• 11:29
  ROI
• 11:29
  per
• 11:30
  dollar
• 11:30
  of
• 11:30
  any
• 11:30
  tool
• 11:31
  on
• 11:31
  the
• 11:31
  list
• 11:31
  because
• 11:32
  review
• 11:32
  tools
• 11:33
  punch
• 11:33
  above
• 11:33
  their
• 11:33
  weight
• 11:34
  when
• 11:34
  you
• 11:34
  are
• 11:34
  using
• 11:34
  generation
• 11:35
  tools
• 11:35
  heavily.
• 11:36
  The
• 11:36
  inversion
• 11:36
  —
• 11:37
  spend
• 11:38
  on
• 11:38
  verification,
• 11:39
  not
• 11:39
  generation
• 11:39
  —
• 11:40
  is
• 11:40
  the
• 11:40
  opposite
• 11:41
  of
• 11:41
  what
• 11:41
  the
• 11:41
  marketing
• 11:42
  tells
• 11:42
  you.
• 11:43
  The
• 11:43
  comments
• 11:43
  thread
• 11:43
  reinforced
• 11:44
  this:
• 11:45
  one
• 11:45
  practitioner
• 11:46
  reported
• 11:46
  that
• 11:46
  a
• 11:46
  review-heavy
• 11:47
  cross-agent
• 11:47
  workflow
• 11:48
  reduced
• 11:48
  a
• 11:48
  15,000-line
• 11:50
  project
• 11:50
  from
• 11:50
  80
• 11:51
  person-days
• 11:51
  (minimal
• 11:52
  AI)
• 11:52
  to
• 11:52
  9
• 11:52
  person-days
• 11:53
  with
• 11:53
  the
• 11:53
  highest
• 11:54
  code
• 11:54
  quality
• 11:55
  of
• 11:55
  any
• 11:55
  approach
• 11:55
  tested.
• 11:56
  The
• 11:57
  capital
• 11:57
  markets
• 11:57
  agree
• 11:58
  that
• 11:58
  the
• 11:58
  implementation
• 11:59
  layer
• 11:59
  matters.
• 12:00
  Anthropic
• 12:00
  and
• 12:01
  Blackstone
• 12:01
  announced
• 12:02
  a
• 12:02
  $1.5
• 12:02
  billion
• 12:03
  deployment
• 12:03
  partnership
• 12:04
  focused
• 12:04
  specifically
• 12:05
  on
• 12:05
  enterprise
• 12:06
  agentic
• 12:06
  workflows.
• 12:08
  OpenAI
• 12:08
  launched
• 12:09
  a
• 12:09
  $10
• 12:09
  billion
• 12:09
  venture
• 12:10
  arm.
• 12:11
  Private
• 12:12
  equity
• 12:12
  firms
• 12:12
  are
• 12:12
  converging
• 12:13
  on
• 12:13
  agentic
• 12:13
  AI
• 12:14
  as
• 12:14
  the
• 12:14
  next
• 12:14
  value-creation
• 12:15
  lever,
• 12:15
  applying
• 12:16
  the
• 12:16
  same
• 12:16
  playbook
• 12:17
  they
• 12:17
  used
• 12:17
  for
• 12:17
  cloud
• 12:17
  migration
• 12:18
  and
• 12:18
  SaaS
• 12:18
  rollups
• 12:19
  a
• 12:19
  decade
• 12:19
  ago.
• 12:20
  The
• 12:20
  pattern
• 12:21
  is
• 12:21
  consistent:
• 12:22
  the
• 12:22
  money
• 12:22
  is
• 12:22
  moving
• 12:23
  from
• 12:23
  model
• 12:23
  training
• 12:23
  to
• 12:23
  model
• 12:24
  deployment
• 12:24
  infrastructure,
• 12:25
  because
• 12:26
  the
• 12:26
  model
• 12:26
  itself
• 12:26
  is
• 12:27
  commoditizing
• 12:27
  while
• 12:28
  the
• 12:28
  integration
• 12:28
  layer
• 12:28
  —
• 12:29
  the
• 12:29
  harness
• 12:30
  —
• 12:30
  is
• 12:30
  where
• 12:31
  defensible
• 12:31
  value
• 12:31
  lives.
• 12:32
  A
• 12:32
  workflow-first
• 12:33
  AI
• 12:33
  investment
• 12:34
  framework
• 12:34
  laid
• 12:34
  out
• 12:34
  five
• 12:35
  levers
• 12:35
  —
• 12:36
  automate,
• 12:36
  build,
• 12:37
  buy,
• 12:37
  hire,
• 12:38
  wait
• 12:38
  —
• 12:39
  and
• 12:39
  made
• 12:39
  the
• 12:39
  case
• 12:39
  that
• 12:40
  the
• 12:40
  "wait"
• 12:40
  option
• 12:41
  carries
• 12:41
  compounding
• 12:41
  opportunity
• 12:42
  cost
• 12:42
  as
• 12:42
  competitors
• 12:43
  capture
• 12:43
  workflow
• 12:44
  data
• 12:44
  that
• 12:44
  feeds
• 12:45
  their
• 12:45
  flywheels.
• 12:46
  The
• 12:46
  Gartner
• 12:46
  line
• 12:46
  that
• 12:46
  more
• 12:47
  than
• 12:47
  30%
• 12:47
  of
• 12:48
  generative
• 12:48
  AI
• 12:48
  projects
• 12:49
  will
• 12:49
  be
• 12:49
  abandoned
• 12:50
  after
• 12:50
  proof-of-concept
• 12:51
  was
• 12:51
  cited
• 12:51
  not
• 12:51
  as
• 12:51
  a
• 12:52
  warning
• 12:52
  against
• 12:52
  investment
• 12:53
  but
• 12:53
  as
• 12:53
  evidence
• 12:54
  that
• 12:54
  the
• 12:54
  implementation
• 12:54
  gap
• 12:55
  is
• 12:55
  where
• 12:55
  the
• 12:55
  real
• 12:55
  competition
• 12:56
  plays
• 12:56
  out.
• 13:00
  TextGen
• 13:00
  (oobabooga)
• 13:01
  shipped
• 13:01
  as
• 13:02
  a
• 13:02
  native
• 13:02
  desktop
• 13:02
  app
• 13:03
  —
• 13:03
  no
• 13:03
  Python
• 13:04
  environment,
• 13:04
  no
• 13:05
  terminal,
• 13:05
  no
• 13:06
  telemetry.
• 13:06
  It
• 13:07
  bundles
• 13:07
  ik
• 13:07
  llama.cpp
• 13:09
  with
• 13:09
  MCP
• 13:09
  support
• 13:10
  and
• 13:10
  tool
• 13:10
  calling,
• 13:11
  runs
• 13:11
  entirely
• 13:11
  offline,
• 13:12
  and
• 13:12
  provides
• 13:13
  a
• 13:13
  clean
• 13:13
  UI
• 13:13
  for
• 13:14
  model
• 13:14
  management,
• 13:14
  chat,
• 13:15
  and
• 13:15
  API
• 13:15
  serving.
• 13:16
  For
• 13:16
  users
• 13:16
  who
• 13:17
  want
• 13:17
  local
• 13:17
  inference
• 13:17
  without
• 13:18
  configuring
• 13:18
  conda
• 13:18
  environments
• 13:19
  or
• 13:19
  debugging
• 13:20
  CUDA
• 13:20
  drivers,
• 13:21
  it
• 13:21
  is
• 13:21
  the
• 13:21
  most
• 13:21
  polished
• 13:22
  option
• 13:22
  currently
• 13:22
  available.
• 13:23
  The
• 13:23
  MCP
• 13:24
  integration
• 13:24
  means
• 13:25
  it
• 13:25
  can
• 13:25
  plug
• 13:25
  into
• 13:25
  agentic
• 13:26
  workflows
• 13:26
  as
• 13:26
  a
• 13:26
  local
• 13:27
  model
• 13:27
  server,
• 13:27
  which
• 13:28
  closes
• 13:28
  a
• 13:28
  gap
• 13:28
  that
• 13:28
  previously
• 13:29
  required
• 13:29
  either
• 13:30
  LM
• 13:30
  Studio
• 13:30
  or
• 13:31
  manual
• 13:31
  llama.cpp
• 13:32
  setup.
• 13:33
  But
• 13:33
  GGUF,
• 13:34
  the
• 13:34
  format
• 13:34
  everything
• 13:35
  local
• 13:35
  runs
• 13:35
  on,
• 13:36
  is
• 13:36
  showing
• 13:36
  its
• 13:36
  age.
• 13:37
  A
• 13:37
  detailed
• 13:38
  analysis
• 13:38
  identified
• 13:39
  four
• 13:39
  critical
• 13:40
  gaps:
• 13:40
  no
• 13:40
  standardized
• 13:41
  tool
• 13:41
  calling
• 13:42
  format
• 13:42
  (every
• 13:42
  runtime
• 13:43
  invents
• 13:43
  its
• 13:43
  own
• 13:44
  chat
• 13:44
  templates
• 13:44
  for
• 13:44
  function
• 13:45
  calling),
• 13:45
  no
• 13:46
  thinking/reasoning
• 13:47
  tag
• 13:47
  convention
• 13:47
  (each
• 13:48
  model
• 13:48
  family
• 13:48
  uses
• 13:49
  different
• 13:49
  special
• 13:49
  tokens),
• 13:50
  no
• 13:50
  multimodal
• 13:51
  bundling
• 13:51
  (vision
• 13:52
  models
• 13:52
  require
• 13:52
  separate
• 13:53
  projection
• 13:53
  files
• 13:53
  with
• 13:54
  no
• 13:54
  standard
• 13:54
  naming),
• 13:55
  and
• 13:55
  no
• 13:55
  feature
• 13:55
  flags
• 13:56
  to
• 13:56
  advertise
• 13:56
  what
• 13:57
  a
• 13:57
  quantized
• 13:57
  model
• 13:58
  actually
• 13:58
  supports.
• 13:59
  When
• 13:59
  a
• 13:59
  user
• 14:00
  downloads
• 14:00
  a
• 14:00
  GGUF,
• 14:01
  they
• 14:01
  cannot
• 14:02
  tell
• 14:02
  from
• 14:02
  the
• 14:02
  file
• 14:02
  alone
• 14:03
  whether
• 14:03
  it
• 14:03
  supports
• 14:03
  tool
• 14:04
  calling,
• 14:04
  what
• 14:05
  chat
• 14:05
  template
• 14:05
  it
• 14:05
  expects,
• 14:06
  or
• 14:06
  whether
• 14:06
  it
• 14:06
  needs
• 14:07
  auxiliary
• 14:07
  files
• 14:08
  to
• 14:08
  function.
• 14:09
  These
• 14:09
  are
• 14:09
  not
• 14:09
  theoretical
• 14:10
  complaints
• 14:10
  —
• 14:11
  they
• 14:11
  are
• 14:11
  the
• 14:11
  reason
• 14:12
  local
• 14:12
  AI
• 14:12
  tooling
• 14:13
  remains
• 14:13
  fragmented
• 14:13
  despite
• 14:14
  individual
• 14:14
  tools
• 14:15
  reaching
• 14:15
  polish.
• 14:16
  The
• 14:17
  format
• 14:17
  is
• 14:17
  the
• 14:17
  bottleneck,
• 14:18
  and
• 14:18
  until
• 14:19
  GGUF
• 14:19
  either
• 14:19
  evolves
• 14:20
  or
• 14:20
  gets
• 14:20
  replaced,
• 14:21
  local
• 14:21
  AI
• 14:22
  will
• 14:22
  keep
• 14:22
  reinventing
• 14:23
  the
• 14:23
  same
• 14:23
  compatibility
• 14:24
  layers.
• 14:25
  On
• 14:25
  the
• 14:25
  voice
• 14:26
  side,
• 14:26
  Supertonic
• 14:27
  from
• 14:27
  Supertone
• 14:28
  shipped
• 14:28
  a
• 14:28
  99-million-parameter
• 14:29
  on-device
• 14:30
  TTS
• 14:30
  engine
• 14:31
  covering
• 14:31
  31
• 14:32
  languages
• 14:32
  with
• 14:32
  ONNX
• 14:33
  Runtime
• 14:33
  inference.
• 14:34
  It
• 14:35
  runs
• 14:35
  on
• 14:35
  Raspberry
• 14:35
  Pi
• 14:36
  and
• 14:36
  e-readers.
• 14:37
  Expression
• 14:38
  tags
• 14:38
  let
• 14:38
  you
• 14:38
  control
• 14:39
  emphasis
• 14:39
  and
• 14:39
  emotion
• 14:40
  without
• 14:40
  prompt
• 14:40
  gymnastics.
• 14:42
  At
• 14:42
  99M
• 14:43
  parameters,
• 14:44
  this
• 14:44
  is
• 14:44
  an
• 14:44
  order
• 14:45
  of
• 14:45
  magnitude
• 14:45
  smaller
• 14:46
  than
• 14:46
  cloud
• 14:46
  TTS
• 14:47
  and
• 14:47
  runs
• 14:47
  entirely
• 14:48
  offline
• 14:48
  —
• 14:49
  the
• 14:49
  kind
• 14:49
  of
• 14:50
  capability
• 14:50
  that
• 14:50
  was
• 14:50
  cloud-only
• 14:51
  a
• 14:51
  year
• 14:51
  ago.
• 14:52
  DramaBox
• 14:53
  from
• 14:53
  Resemble
• 14:53
  AI
• 14:54
  claims
• 14:54
  the
• 14:54
  most
• 14:54
  expressive
• 14:55
  voice
• 14:55
  model
• 14:55
  yet,
• 14:56
  built
• 14:56
  on
• 14:56
  LTX
• 14:57
  2.3.
• 14:58
  GenCAD
• 14:59
  demonstrated
• 15:00
  image-conditional
• 15:00
  parametric
• 15:01
  CAD
• 15:01
  generation
• 15:02
  —
• 15:02
  feed
• 15:03
  it
• 15:03
  a
• 15:03
  photo
• 15:03
  of
• 15:03
  a
• 15:03
  mechanical
• 15:04
  part
• 15:04
  and
• 15:04
  get
• 15:04
  editable
• 15:05
  CAD
• 15:05
  geometry
• 15:05
  —
• 15:06
  crossing
• 15:07
  AI
• 15:07
  into
• 15:07
  physical
• 15:07
  manufacturing
• 15:08
  design
• 15:09
  where
• 15:09
  the
• 15:09
  stakes
• 15:09
  include
• 15:09
  structural
• 15:10
  integrity,
• 15:11
  not
• 15:11
  just
• 15:11
  aesthetics.