Latest / AGI Dreams – Open, Uncensored, & Local - AI News Digest / AGI Dreams Podcast – March 06, 2026

Transcript

AGI Dreams Podcast – March 06, 2026 · AGI Dreams – Open, Uncensored, & Local - AI News Digest

• 0:00
  Apple
• 0:00
  dropped
• 0:00
  the
• 0:00
  M5
• 0:01
  Pro
• 0:01
  and
• 0:01
  M5
• 0:02
  Max
• 0:02
  MacBook
• 0:03
  Pros
• 0:03
  this
• 0:03
  week,
• 0:03
  and
• 0:04
  the
• 0:04
  spec
• 0:04
  sheet
• 0:04
  reads
• 0:04
  like
• 0:05
  a
• 0:05
  declaration
• 0:05
  of
• 0:05
  intent.
• 0:06
  The
• 0:06
  new
• 0:07
  chips
• 0:07
  use
• 0:07
  a
• 0:07
  "Fusion
• 0:08
  Architecture"
• 0:08
  that
• 0:08
  combines
• 0:09
  two
• 0:09
  dies
• 0:09
  into
• 0:10
  a
• 0:10
  single
• 0:10
  SoC,
• 0:11
  embedding
• 0:11
  a
• 0:11
  Neural
• 0:12
  Accelerator
• 0:12
  in
• 0:13
  every
• 0:13
  GPU
• 0:13
  core.
• 0:14
  The
• 0:14
  results:
• 0:15
  up
• 0:15
  to
• 0:15
  4x
• 0:16
  faster
• 0:16
  LLM
• 0:17
  prompt
• 0:17
  processing
• 0:18
  versus
• 0:18
  M4
• 0:19
  Pro/Max,
• 0:20
  up
• 0:20
  to
• 0:21
  8x
• 0:21
  faster
• 0:21
  AI
• 0:22
  image
• 0:22
  generation
• 0:22
  compared
• 0:23
  to
• 0:23
  M1
• 0:23
  models,
• 0:24
  and
• 0:24
  128GB
• 0:25
  of
• 0:26
  unified
• 0:26
  memory
• 0:26
  on
• 0:26
  the
• 0:27
  M5
• 0:27
  Max
• 0:27
  with
• 0:27
  614GB/s
• 0:30
  bandwidth.
• 0:31
  SSD
• 0:31
  speeds
• 0:31
  hit
• 0:32
  14.5GB/s
• 0:33
  —
• 0:35
  roughly
• 0:35
  2x
• 0:36
  the
• 0:36
  prior
• 0:36
  generation
• 0:36
  —
• 0:37
  and
• 0:37
  base
• 0:37
  storage
• 0:37
  jumps
• 0:38
  to
• 0:38
  1TB
• 0:38
  for
• 0:39
  M5
• 0:39
  Pro,
• 0:39
  2TB
• 0:40
  for
• 0:40
  M5
• 0:40
  Max.
• 0:41
  The
• 0:41
  18-core
• 0:42
  CPU
• 0:42
  features
• 0:43
  6
• 0:43
  "super
• 0:44
  cores"
• 0:45
  Apple
• 0:45
  claims
• 0:45
  are
• 0:45
  the
• 0:46
  world's
• 0:46
  fastest
• 0:46
  CPU
• 0:47
  cores,
• 0:48
  alongside
• 0:48
  12
• 0:48
  performance
• 0:49
  cores
• 0:49
  tuned
• 0:50
  for
• 0:50
  multithreaded
• 0:50
  workloads.
• 0:52
  Apple's
• 0:52
  pitch
• 0:52
  is
• 0:52
  explicit:
• 0:53
  run
• 0:53
  advanced
• 0:54
  LLMs
• 0:54
  on-device,
• 0:55
  train
• 0:55
  custom
• 0:56
  models
• 0:56
  locally,
• 0:57
  do
• 0:57
  it
• 0:57
  all
• 0:57
  on
• 0:57
  battery
• 0:58
  for
• 0:58
  up
• 0:58
  to
• 0:58
  24
• 0:58
  hours.
• 0:59
  The
• 0:59
  hardware
• 1:00
  is
• 1:00
  only
• 1:00
  half
• 1:00
  the
• 1:00
  move.
• 1:01
  Apple
• 1:02
  simultaneously
• 1:02
  released
• 1:03
  the
• 1:03
  Foundation
• 1:03
  Models
• 1:03
  SDK
• 1:04
  for
• 1:04
  Python
• 1:04
  —
• 1:05
  an
• 1:05
  open-source
• 1:06
  package
• 1:06
  that
• 1:06
  provides
• 1:07
  Python
• 1:07
  bindings
• 1:08
  to
• 1:08
  the
• 1:08
  on-device
• 1:08
  foundation
• 1:09
  model
• 1:09
  powering
• 1:10
  Apple
• 1:10
  Intelligence
• 1:10
  on
• 1:10
  macOS
• 1:11
  26.
• 1:12
  The
• 1:12
  SDK
• 1:13
  lets
• 1:13
  developers
• 1:14
  run
• 1:14
  batch
• 1:14
  inference,
• 1:15
  stream
• 1:15
  text
• 1:15
  generation,
• 1:16
  and
• 1:16
  use
• 1:16
  guided
• 1:17
  generation
• 1:17
  with
• 1:17
  structured
• 1:18
  output
• 1:18
  schemas,
• 1:19
  all
• 1:19
  hitting
• 1:20
  the
• 1:20
  system
• 1:20
  model
• 1:20
  locally
• 1:21
  with
• 1:21
  zero
• 1:21
  cloud
• 1:21
  calls.
• 1:22
  It
• 1:22
  requires
• 1:23
  macOS
• 1:23
  26
• 1:24
  and
• 1:24
  Xcode
• 1:25
  26,
• 1:26
  which
• 1:26
  means
• 1:26
  it's
• 1:26
  tightly
• 1:27
  coupled
• 1:27
  to
• 1:27
  the
• 1:27
  new
• 1:27
  OS
• 1:28
  generation,
• 1:29
  but
• 1:29
  for
• 1:29
  the
• 1:29
  Python-heavy
• 1:30
  ML
• 1:30
  community,
• 1:31
  this
• 1:31
  is
• 1:31
  the
• 1:31
  first
• 1:31
  official
• 1:32
  on-ramp
• 1:32
  to
• 1:32
  Apple's
• 1:33
  on-device
• 1:33
  inference
• 1:33
  stack
• 1:34
  without
• 1:34
  writing
• 1:34
  Swift.
• 1:35
  The
• 1:36
  strategic
• 1:36
  narrative
• 1:36
  writes
• 1:37
  itself,
• 1:37
  and
• 1:37
  commentators
• 1:38
  aren't
• 1:38
  being
• 1:38
  subtle
• 1:39
  about
• 1:39
  it.
• 1:40
  While
• 1:40
  the
• 1:40
  rest
• 1:40
  of
• 1:40
  the
• 1:40
  industry
• 1:41
  burned
• 1:41
  through
• 1:41
  $1.4
• 1:42
  trillion-plus
• 1:43
  training
• 1:43
  frontier
• 1:43
  models,
• 1:44
  Apple
• 1:45
  partnered
• 1:45
  with
• 1:45
  Google
• 1:45
  to
• 1:45
  power
• 1:46
  Siri
• 1:46
  with
• 1:46
  Gemini,
• 1:47
  paying
• 1:47
  roughly
• 1:48
  $1
• 1:48
  billion
• 1:48
  per
• 1:49
  year
• 1:49
  to
• 1:49
  integrate
• 1:49
  the
• 1:49
  best
• 1:49
  external
• 1:50
  model
• 1:50
  while
• 1:51
  investing
• 1:51
  its
• 1:51
  silicon
• 1:51
  budget
• 1:52
  into
• 1:52
  making
• 1:52
  on-device
• 1:53
  inference
• 1:53
  competitive.
• 1:54
  The
• 1:54
  argument:
• 1:55
  the
• 1:55
  AI
• 1:56
  era
• 1:56
  won't
• 1:56
  be
• 1:56
  decided
• 1:57
  by
• 1:57
  who
• 1:57
  builds
• 1:57
  the
• 1:57
  smartest
• 1:58
  model,
• 1:58
  but
• 1:58
  by
• 1:58
  who
• 1:59
  owns
• 1:59
  the
• 1:59
  device
• 1:59
  that
• 1:59
  model
• 2:00
  runs
• 2:00
  on.
• 2:01
  Apple
• 2:01
  flips
• 2:02
  a
• 2:02
  switch
• 2:02
  and
• 2:02
  reaches
• 2:02
  2.5
• 2:03
  billion
• 2:03
  devices.
• 2:05
  Competitors
• 2:05
  chase
• 2:05
  subscriptions
• 2:06
  at
• 2:06
  $20–$200/month;
• 2:08
  Apple
• 2:09
  runs
• 2:09
  inference
• 2:09
  on-device
• 2:09
  —
• 2:11
  faster,
• 2:11
  cheaper,
• 2:12
  private.
• 2:13
  Whether
• 2:13
  this
• 2:14
  bet
• 2:14
  pays
• 2:14
  off
• 2:14
  depends
• 2:15
  on
• 2:15
  whether
• 2:15
  on-device
• 2:16
  models
• 2:16
  can
• 2:16
  close
• 2:16
  the
• 2:17
  capability
• 2:17
  gap
• 2:17
  with
• 2:17
  cloud
• 2:18
  frontier
• 2:18
  models
• 2:19
  fast
• 2:19
  enough,
• 2:20
  but
• 2:20
  the
• 2:20
  distribution
• 2:21
  moat
• 2:21
  is
• 2:21
  real.
• 2:26
  Sergej
• 2:27
  Epp,
• 2:27
  CISO
• 2:28
  of
• 2:28
  Sysdig,
• 2:28
  launched
• 2:28
  the
• 2:28
  Zero
• 2:29
  Day
• 2:29
  Clock
• 2:29
  this
• 2:29
  week
• 2:30
  —
• 2:30
  a
• 2:30
  live
• 2:31
  dashboard
• 2:31
  tracking
• 2:31
  the
• 2:32
  collapse
• 2:32
  of
• 2:32
  time-to-exploit
• 2:33
  (TTE)
• 2:33
  from
• 2:33
  CVE
• 2:34
  disclosure
• 2:34
  to
• 2:34
  confirmed
• 2:35
  exploitation
• 2:35
  in
• 2:35
  the
• 2:35
  wild,
• 2:36
  built
• 2:37
  on
• 2:37
  3,515
• 2:38
  CVE-exploit
• 2:39
  pairs
• 2:39
  from
• 2:40
  CISA
• 2:40
  KEV,
• 2:40
  VulnCheck
• 2:41
  KEV,
• 2:41
  and
• 2:41
  XDB.
• 2:42
  The
• 2:42
  trajectory
• 2:43
  is
• 2:43
  brutal.
• 2:44
  In
• 2:44
  2018,
• 2:45
  the
• 2:45
  median
• 2:45
  TTE
• 2:46
  was
• 2:46
  771
• 2:47
  days.
• 2:48
  By
• 2:48
  2021:
• 2:49
  84
• 2:50
  days.
• 2:51
  By
• 2:51
  2023:
• 2:52
  6
• 2:52
  days.
• 2:53
  By
• 2:53
  2024:
• 2:54
  4
• 2:55
  hours.
• 2:56
  In
• 2:56
  2025,
• 2:57
  the
• 2:57
  majority
• 2:58
  of
• 2:58
  exploited
• 2:58
  vulnerabilities
• 2:59
  were
• 2:59
  weaponized
• 3:00
  before
• 3:00
  they
• 3:00
  were
• 3:00
  publicly
• 3:01
  disclosed.
• 3:02
  The
• 3:02
  1-day
• 3:03
  and
• 3:03
  1-hour
• 3:03
  marks
• 3:04
  are
• 3:04
  projected
• 3:04
  for
• 3:04
  2026,
• 3:05
  with
• 3:05
  1
• 3:06
  minute
• 3:06
  projected
• 3:06
  for
• 3:06
  2028.
• 3:08
  Organizations
• 3:09
  need
• 3:09
  an
• 3:09
  average
• 3:09
  of
• 3:09
  20
• 3:09
  days
• 3:10
  to
• 3:10
  test
• 3:10
  and
• 3:10
  deploy
• 3:11
  a
• 3:11
  patch,
• 3:11
  but
• 3:11
  AI
• 3:12
  can
• 3:12
  now
• 3:12
  reverse-engineer
• 3:13
  a
• 3:13
  patch,
• 3:13
  identify
• 3:14
  the
• 3:14
  vulnerability,
• 3:15
  and
• 3:15
  generate
• 3:15
  a
• 3:15
  working
• 3:15
  exploit
• 3:16
  in
• 3:16
  minutes.
• 3:17
  The
• 3:17
  act
• 3:17
  of
• 3:17
  fixing
• 3:17
  a
• 3:17
  vulnerability
• 3:18
  now
• 3:18
  accelerates
• 3:19
  its
• 3:19
  exploitation.
• 3:20
  Epp
• 3:20
  introduces
• 3:21
  "Verifier's
• 3:22
  Law"
• 3:22
  to
• 3:22
  explain
• 3:22
  the
• 3:22
  asymmetry:
• 3:23
  in
• 3:23
  offense,
• 3:24
  feedback
• 3:24
  is
• 3:24
  binary
• 3:25
  and
• 3:25
  instant
• 3:25
  (did
• 3:25
  the
• 3:26
  exploit
• 3:26
  work?),
• 3:26
  so
• 3:27
  AI
• 3:27
  learns
• 3:27
  at
• 3:27
  machine
• 3:28
  speed.
• 3:28
  In
• 3:29
  defense,
• 3:29
  feedback
• 3:30
  is
• 3:30
  ambiguous,
• 3:30
  slow,
• 3:31
  and
• 3:31
  expensive.
• 3:32
  The
• 3:32
  annual
• 3:32
  CVE
• 3:33
  count
• 3:33
  hit
• 3:33
  48,000
• 3:34
  in
• 3:35
  2025
• 3:35
  —
• 3:36
  a
• 3:36
  520%
• 3:37
  rise
• 3:38
  since
• 3:38
  2016
• 3:38
  —
• 3:39
  and
• 3:39
  67.2%
• 3:41
  of
• 3:41
  exploited
• 3:41
  CVEs
• 3:42
  in
• 3:42
  2026
• 3:43
  are
• 3:43
  zero-days,
• 3:44
  up
• 3:44
  from
• 3:44
  16.1%
• 3:45
  in
• 3:45
  2018.
• 3:46
  Monthly
• 3:47
  patch
• 3:47
  cycles
• 3:47
  aren't
• 3:47
  inadequate;
• 3:48
  they're
• 3:48
  irrelevant.
• 3:50
  A
• 3:50
  companion
• 3:50
  story
• 3:50
  from
• 3:50
  Thai
• 3:51
  Duong
• 3:51
  at
• 3:51
  Calif
• 3:52
  puts
• 3:52
  a
• 3:52
  human
• 3:52
  face
• 3:52
  on
• 3:52
  the
• 3:53
  offensive
• 3:53
  side.
• 3:54
  A
• 3:54
  self-taught
• 3:55
  hacker
• 3:55
  from
• 3:55
  Cần
• 3:55
  Thơ,
• 3:56
  Vietnam
• 3:56
  —
• 3:57
  no
• 3:57
  degree,
• 3:57
  no
• 3:58
  bootcamp
• 3:58
  certificate,
• 3:59
  dropped
• 3:59
  out
• 3:59
  of
• 3:59
  university
• 4:00
  before
• 4:00
  20
• 4:00
  —
• 4:01
  found
• 4:01
  six
• 4:01
  vulnerabilities
• 4:02
  in
• 4:02
  Google's
• 4:03
  kernelCTF
• 4:03
  and
• 4:03
  exploited
• 4:04
  four,
• 4:04
  including
• 4:05
  a
• 4:05
  20-year-old
• 4:05
  Linux
• 4:06
  kernel
• 4:06
  bug
• 4:06
  (CVE-2025-38617).
• 4:10
  The
• 4:10
  bug
• 4:11
  itself
• 4:11
  was
• 4:11
  simple,
• 4:11
  which
• 4:12
  is
• 4:12
  the
• 4:12
  terrifying
• 4:12
  part:
• 4:13
  it
• 4:13
  survived
• 4:13
  two
• 4:13
  decades
• 4:14
  because
• 4:14
  it
• 4:14
  takes
• 4:14
  a
• 4:14
  particular
• 4:15
  kind
• 4:15
  of
• 4:15
  mind
• 4:15
  to
• 4:15
  notice
• 4:16
  it.
• 4:17
  The
• 4:17
  exploit,
• 4:17
  however,
• 4:17
  was
• 4:18
  "fine
• 4:18
  art"
• 4:18
  —
• 4:19
  Thai
• 4:19
  Duong
• 4:19
  says
• 4:19
  it
• 4:19
  took
• 4:19
  him
• 4:20
  four
• 4:20
  days
• 4:20
  and
• 4:20
  "an
• 4:20
  embarrassing
• 4:21
  number
• 4:21
  of
• 4:21
  tokens"
• 4:21
  to
• 4:21
  understand
• 4:22
  how
• 4:22
  it
• 4:22
  worked.
• 4:23
  Calif
• 4:23
  published
• 4:23
  a
• 4:23
  step-by-step
• 4:24
  exploitation
• 4:25
  guide.
• 4:25
  The
• 4:25
  story
• 4:26
  is
• 4:26
  both
• 4:26
  an
• 4:26
  inspiring
• 4:27
  talent
• 4:27
  narrative
• 4:27
  and
• 4:27
  a
• 4:27
  sobering
• 4:28
  reminder
• 4:28
  that
• 4:28
  the
• 4:29
  talent
• 4:29
  pool
• 4:29
  for
• 4:29
  offensive
• 4:29
  security
• 4:30
  is
• 4:30
  global,
• 4:30
  deep,
• 4:31
  and
• 4:31
  not
• 4:31
  always
• 4:31
  visible
• 4:32
  to
• 4:32
  the
• 4:32
  industry's
• 4:32
  traditional
• 4:33
  hiring
• 4:33
  pipelines.
• 4:34
  The
• 4:34
  [un]prompted
• 4:35
  2026
• 4:35
  conference
• 4:36
  in
• 4:36
  San
• 4:36
  Francisco
• 4:37
  crystallized
• 4:37
  how
• 4:38
  fast
• 4:38
  the
• 4:38
  offensive-defensive
• 4:39
  timeline
• 4:39
  is
• 4:39
  compressing.
• 4:41
  Nicholas
• 4:41
  Carlini
• 4:41
  from
• 4:42
  Anthropic
• 4:42
  showed
• 4:42
  LLMs
• 4:43
  autonomously
• 4:44
  finding
• 4:45
  and
• 4:45
  exploiting
• 4:45
  zero-day
• 4:46
  vulnerabilities
• 4:46
  in
• 4:47
  production
• 4:47
  software
• 4:47
  —
• 4:48
  a
• 4:48
  heap
• 4:49
  buffer
• 4:49
  overflow
• 4:49
  in
• 4:49
  the
• 4:50
  Linux
• 4:50
  kernel
• 4:50
  hiding
• 4:50
  since
• 4:51
  2003,
• 4:52
  the
• 4:52
  first
• 4:52
  critical
• 4:53
  CVE
• 4:53
  in
• 4:54
  Ghost
• 4:54
  CMS
• 4:54
  —
• 4:55
  declaring
• 4:55
  "these
• 4:56
  current
• 4:56
  models
• 4:56
  are
• 4:56
  better
• 4:57
  vulnerability
• 4:57
  researchers
• 4:58
  than
• 4:58
  I
• 4:58
  am."
• 4:59
  Google's
• 4:59
  Heather
• 5:00
  Adkins
• 5:00
  stated
• 5:00
  a
• 5:01
  near-term
• 5:01
  goal:
• 5:02
  shipping
• 5:02
  relatively
• 5:03
  bug-free
• 5:03
  code
• 5:04
  within
• 5:04
  two
• 5:04
  years,
• 5:05
  backed
• 5:05
  by
• 5:05
  Big
• 5:06
  Sleep
• 5:06
  (zero
• 5:06
  false
• 5:06
  positives
• 5:07
  on
• 5:07
  memory
• 5:07
  safety
• 5:08
  bugs)
• 5:09
  and
• 5:09
  CodeMender
• 5:09
  (178
• 5:10
  autonomous
• 5:11
  patches).
• 5:12
  Trail
• 5:12
  of
• 5:12
  Bits
• 5:13
  went
• 5:13
  from
• 5:13
  15
• 5:13
  to
• 5:14
  200
• 5:14
  bugs
• 5:14
  per
• 5:15
  week
• 5:15
  per
• 5:15
  engineer
• 5:15
  using
• 5:16
  AI
• 5:16
  agent
• 5:16
  fleets.
• 5:18
  Derek
• 5:18
  Chen
• 5:18
  reported
• 5:19
  AI-powered
• 5:19
  vulnerability
• 5:20
  discovery
• 5:20
  at
• 5:20
  61
• 5:21
  cents
• 5:21
  per
• 5:22
  finding,
• 5:22
  with
• 5:23
  60+
• 5:23
  CVEs
• 5:24
  submitted
• 5:24
  and
• 5:24
  3,000+
• 5:26
  pending
• 5:26
  review.
• 5:28
  Day
• 5:28
  two
• 5:28
  of
• 5:28
  [un]prompted
• 5:29
  added
• 5:29
  equally
• 5:29
  sharp
• 5:30
  data
• 5:30
  points.
• 5:31
  Rob
• 5:32
  Lee
• 5:32
  from
• 5:32
  SANS
• 5:32
  compressed
• 5:33
  a
• 5:33
  3-day
• 5:33
  intrusion
• 5:34
  investigation
• 5:35
  into
• 5:35
  14
• 5:35
  minutes
• 5:36
  using
• 5:36
  Claude
• 5:36
  with
• 5:36
  a
• 5:36
  forensic-skills
• 5:37
  CLAUDE.md
• 5:38
  file.
• 5:39
  A
• 5:40
  real-world
• 5:40
  AI-assisted
• 5:41
  AWS
• 5:42
  attack
• 5:42
  went
• 5:42
  from
• 5:42
  stolen
• 5:43
  credentials
• 5:43
  to
• 5:43
  full
• 5:44
  admin
• 5:44
  in
• 5:44
  8
• 5:45
  minutes
• 5:45
  —
• 5:46
  but
• 5:46
  the
• 5:46
  AI
• 5:46
  "accent"
• 5:47
  (hallucinated
• 5:48
  GitHub
• 5:48
  repos,
• 5:49
  training-set
• 5:50
  sample
• 5:50
  account
• 5:50
  IDs,
• 5:51
  a
• 5:51
  GPU
• 5:52
  cluster
• 5:52
  named
• 5:52
  "steven
• 5:53
  gpu
• 5:53
  monster")
• 5:54
  made
• 5:54
  it
• 5:54
  the
• 5:54
  "noisiest
• 5:55
  attack
• 5:55
  we've
• 5:55
  ever
• 5:56
  seen."
• 5:57
  Researchers
• 5:58
  found
• 5:58
  37
• 5:58
  vulnerabilities
• 5:59
  across
• 5:59
  15+
• 6:00
  AI
• 6:01
  IDE
• 6:01
  vendors,
• 6:02
  including
• 6:02
  a
• 6:02
  zero-click
• 6:03
  MCP
• 6:03
  autoload
• 6:04
  attack
• 6:04
  in
• 6:04
  Codex
• 6:05
  that
• 6:05
  spawns
• 6:06
  a
• 6:06
  reverse
• 6:06
  shell
• 6:06
  on
• 6:07
  workspace
• 6:07
  open.
• 6:08
  Snap
• 6:09
  presented
• 6:09
  "Tenu
• 6:09
  warrants"
• 6:10
  —
• 6:11
  capability-based
• 6:12
  tokens
• 6:12
  that
• 6:12
  reduced
• 6:13
  successful
• 6:13
  AI
• 6:13
  agent
• 6:14
  attack
• 6:14
  surface
• 6:14
  from
• 6:15
  90%
• 6:15
  to
• 6:15
  0%
• 6:16
  by
• 6:16
  constraining
• 6:17
  agent
• 6:17
  permissions
• 6:17
  at
• 6:18
  execution
• 6:18
  time
• 6:19
  even
• 6:19
  if
• 6:19
  prompt-injected.
• 6:20
  Johann
• 6:21
  Rehberger
• 6:21
  coined
• 6:22
  "promptware"
• 6:23
  as
• 6:23
  the
• 6:23
  successor
• 6:23
  to
• 6:23
  malware:
• 6:24
  multi-stage,
• 6:25
  persistent
• 6:26
  instruction
• 6:26
  sets
• 6:27
  operating
• 6:27
  above
• 6:28
  the
• 6:28
  OS
• 6:28
  layer
• 6:28
  via
• 6:29
  prompt
• 6:29
  injection.
• 6:33
  NeuroSploit
• 6:34
  v3
• 6:34
  is
• 6:34
  an
• 6:34
  open-source
• 6:35
  AI-powered
• 6:36
  penetration
• 6:36
  testing
• 6:37
  platform
• 6:37
  that
• 6:37
  illustrates
• 6:38
  how
• 6:38
  far
• 6:38
  autonomous
• 6:39
  offensive
• 6:39
  tooling
• 6:39
  has
• 6:39
  come.
• 6:40
  It
• 6:40
  covers
• 6:40
  100
• 6:41
  vulnerability
• 6:41
  types
• 6:41
  across
• 6:42
  10
• 6:42
  categories,
• 6:43
  runs
• 6:43
  a
• 6:43
  3-stream
• 6:43
  parallel
• 6:44
  architecture
• 6:44
  (recon,
• 6:45
  junior
• 6:45
  tester,
• 6:45
  tool
• 6:45
  runner)
• 6:46
  with
• 6:46
  per-scan
• 6:46
  isolated
• 6:47
  Kali
• 6:47
  Linux
• 6:47
  containers,
• 6:48
  and
• 6:48
  includes
• 6:48
  a
• 6:48
  multi-layered
• 6:49
  anti-hallucination
• 6:49
  pipeline:
• 6:50
  negative
• 6:51
  controls
• 6:51
  send
• 6:51
  benign
• 6:51
  requests
• 6:52
  as
• 6:52
  baselines,
• 6:53
  25+
• 6:54
  per-vulnerability
• 6:54
  proof-of-execution
• 6:55
  methods
• 6:55
  verify
• 6:56
  findings,
• 6:56
  a
• 6:56
  confidence
• 6:57
  scorer
• 6:57
  rates
• 6:57
  0–100,
• 6:59
  and
• 6:59
  a
• 6:59
  validation
• 7:00
  judge
• 7:00
  serves
• 7:00
  as
• 7:01
  final
• 7:01
  authority.
• 7:02
  The
• 7:02
  payload
• 7:02
  engine
• 7:02
  ships
• 7:03
  526
• 7:04
  payloads
• 7:04
  across
• 7:04
  95
• 7:05
  libraries,
• 7:05
  with
• 7:05
  WAF-adaptive
• 7:06
  transformation
• 7:06
  covering
• 7:07
  16
• 7:07
  WAF
• 7:07
  signatures
• 7:08
  and
• 7:08
  12
• 7:08
  bypass
• 7:09
  techniques.
• 7:09
  Exploit
• 7:10
  chaining
• 7:10
  is
• 7:10
  automated
• 7:11
  —
• 7:11
  SSRF
• 7:12
  to
• 7:12
  internal
• 7:12
  access,
• 7:13
  SQLi
• 7:13
  to
• 7:13
  DB-specific
• 7:14
  payloads,
• 7:15
  LFI
• 7:15
  to
• 7:15
  config
• 7:16
  extraction.
• 7:17
  It's
• 7:17
  multi-provider
• 7:17
  (Claude,
• 7:18
  GPT,
• 7:19
  Gemini,
• 7:19
  Ollama,
• 7:20
  LMStudio)
• 7:21
  and
• 7:21
  includes
• 7:21
  an
• 7:21
  MCP
• 7:21
  server
• 7:22
  with
• 7:22
  12
• 7:22
  tools.
• 7:23
  The
• 7:23
  maturity
• 7:24
  here
• 7:24
  is
• 7:24
  notable:
• 7:25
  this
• 7:25
  isn't
• 7:25
  a
• 7:25
  proof-of-concept
• 7:26
  wrapper
• 7:26
  around
• 7:26
  an
• 7:27
  LLM,
• 7:27
  it's
• 7:28
  a
• 7:28
  full
• 7:28
  autonomous
• 7:28
  pentesting
• 7:29
  stack
• 7:29
  with
• 7:29
  cross-scan
• 7:30
  learning
• 7:30
  and
• 7:30
  adaptive
• 7:30
  strategy.
• 7:32
  On
• 7:32
  the
• 7:32
  theoretical
• 7:33
  side,
• 7:33
  Pete
• 7:33
  Herzog
• 7:34
  published
• 7:34
  what
• 7:34
  he
• 7:34
  calls
• 7:35
  the
• 7:35
  "Grand
• 7:35
  Unified
• 7:36
  Security
• 7:36
  Theory"
• 7:36
  (GUST)
• 7:36
  —
• 7:37
  a
• 7:37
  mapping
• 7:37
  of
• 7:38
  240
• 7:38
  interconnected
• 7:39
  security
• 7:39
  elements
• 7:40
  across
• 7:40
  four
• 7:40
  dimensions,
• 7:41
  derived
• 7:41
  from
• 7:42
  quantum
• 7:42
  information
• 7:42
  properties
• 7:43
  applied
• 7:43
  to
• 7:43
  Newtonian
• 7:44
  spacetime.
• 7:45
  The
• 7:45
  framework
• 7:46
  identifies
• 7:46
  16
• 7:47
  categories
• 7:47
  of
• 7:47
  measurable
• 7:48
  security
• 7:48
  reality,
• 7:49
  each
• 7:49
  containing
• 7:50
  15
• 7:50
  discrete
• 7:51
  elements
• 7:51
  across
• 7:51
  three
• 7:51
  operational
• 7:52
  phases
• 7:52
  (before,
• 7:53
  during,
• 7:54
  after
• 7:54
  an
• 7:54
  interaction),
• 7:55
  organized
• 7:56
  into
• 7:56
  positive
• 7:56
  categories
• 7:57
  (Security
• 7:57
  Controls,
• 7:58
  Trust
• 7:58
  Properties)
• 7:59
  and
• 7:59
  negative
• 7:59
  categories
• 8:00
  (Exploitation,
• 8:00
  Privacy).
• 8:02
  Herzog's
• 8:02
  core
• 8:02
  claim:
• 8:03
  a
• 8:03
  typical
• 8:04
  security
• 8:04
  checklist
• 8:05
  covers
• 8:05
  20–40
• 8:06
  of
• 8:07
  these
• 8:07
  240
• 8:08
  elements,
• 8:08
  and
• 8:09
  even
• 8:09
  the
• 8:09
  most
• 8:09
  mature
• 8:09
  compliance
• 8:10
  framework
• 8:10
  touches
• 8:11
  around
• 8:11
  80.
• 8:12
  The
• 8:12
  practical
• 8:13
  implication
• 8:13
  isn't
• 8:13
  that
• 8:14
  you
• 8:14
  need
• 8:14
  to
• 8:14
  master
• 8:14
  all
• 8:15
  240
• 8:15
  —
• 8:16
  it's
• 8:16
  that
• 8:16
  you
• 8:16
  don't
• 8:17
  know
• 8:17
  which
• 8:17
  gaps
• 8:17
  are
• 8:17
  irrelevant
• 8:18
  without
• 8:18
  mapping
• 8:18
  the
• 8:19
  full
• 8:19
  territory
• 8:19
  first.
• 8:20
  Whether
• 8:21
  GUST
• 8:21
  survives
• 8:21
  contact
• 8:22
  with
• 8:22
  real-world
• 8:23
  practice
• 8:23
  remains
• 8:23
  to
• 8:23
  be
• 8:24
  seen,
• 8:24
  but
• 8:25
  the
• 8:25
  framing
• 8:25
  of
• 8:25
  security-as-incomplete-map
• 8:27
  is
• 8:27
  a
• 8:27
  useful
• 8:27
  corrective
• 8:28
  to
• 8:28
  "good
• 8:28
  enough"
• 8:28
  compliance
• 8:29
  thinking.
• 8:34
  Liquid
• 8:34
  AI
• 8:35
  released
• 8:35
  LFM2-24B-A2B
• 8:38
  alongside
• 8:39
  LocalCowork,
• 8:39
  an
• 8:39
  open-source
• 8:40
  desktop
• 8:40
  agent
• 8:41
  that
• 8:41
  runs
• 8:41
  entirely
• 8:41
  on-device
• 8:42
  with
• 8:42
  no
• 8:42
  cloud
• 8:42
  dependency.
• 8:43
  Tested
• 8:43
  against
• 8:44
  67
• 8:44
  tools
• 8:44
  across
• 8:45
  13
• 8:45
  MCP
• 8:45
  servers
• 8:46
  on
• 8:46
  a
• 8:46
  MacBook
• 8:46
  Pro
• 8:47
  with
• 8:47
  36GB
• 8:48
  unified
• 8:48
  memory,
• 8:49
  LFM2
• 8:49
  averaged
• 8:50
  sub-400ms
• 8:51
  tool
• 8:51
  selection
• 8:52
  while
• 8:52
  fitting
• 8:52
  in
• 8:52
  roughly
• 8:53
  14.5GB
• 8:55
  of
• 8:55
  memory.
• 8:56
  Single-step
• 8:56
  tool
• 8:57
  selection
• 8:57
  accuracy
• 8:57
  hit
• 8:58
  80%
• 8:58
  across
• 8:58
  100
• 8:59
  prompts,
• 8:59
  which
• 8:59
  sounds
• 9:00
  modest
• 9:00
  until
• 9:00
  you
• 9:00
  consider
• 9:01
  the
• 9:01
  interaction
• 9:01
  model:
• 9:02
  the
• 9:02
  agent
• 9:03
  proposes
• 9:03
  a
• 9:03
  tool,
• 9:03
  the
• 9:04
  user
• 9:04
  confirms
• 9:04
  or
• 9:04
  corrects,
• 9:05
  and
• 9:05
  when
• 9:05
  the
• 9:05
  loop
• 9:06
  is
• 9:06
  fast
• 9:06
  enough,
• 9:06
  even
• 9:06
  imperfect
• 9:07
  accuracy
• 9:07
  becomes
• 9:08
  usable
• 9:08
  because
• 9:08
  correction
• 9:09
  costs
• 9:09
  are
• 9:09
  low.
• 9:10
  Multi-step
• 9:11
  chain
• 9:11
  completion
• 9:11
  was
• 9:12
  weaker
• 9:12
  at
• 9:12
  26%,
• 9:13
  which
• 9:13
  Liquid
• 9:13
  acknowledges
• 9:14
  —
• 9:15
  LFM2
• 9:15
  works
• 9:15
  best
• 9:16
  as
• 9:16
  a
• 9:16
  fast
• 9:16
  dispatcher
• 9:17
  in
• 9:17
  a
• 9:17
  guided
• 9:17
  loop,
• 9:17
  not
• 9:18
  an
• 9:18
  autonomous
• 9:18
  autopilot.
• 9:19
  The
• 9:20
  model
• 9:20
  has
• 9:20
  been
• 9:20
  trained
• 9:20
  on
• 9:20
  17T
• 9:21
  tokens
• 9:21
  with
• 9:22
  pre-training
• 9:22
  still
• 9:22
  running;
• 9:23
  an
• 9:23
  LFM2.5
• 9:25
  release
• 9:26
  with
• 9:26
  post-training
• 9:26
  and
• 9:26
  reinforcement
• 9:27
  learning
• 9:27
  is
• 9:27
  expected.
• 9:28
  For
• 9:28
  privacy-sensitive
• 9:29
  and
• 9:29
  regulated
• 9:30
  workloads
• 9:30
  where
• 9:31
  no
• 9:31
  data
• 9:31
  can
• 9:31
  leave
• 9:31
  the
• 9:31
  device,
• 9:32
  this
• 9:32
  is
• 9:32
  a
• 9:32
  meaningful
• 9:33
  entry.
• 9:34
  Claude
• 9:34
  Cowork
• 9:34
  is
• 9:35
  getting
• 9:35
  the
• 9:35
  COO
• 9:35
  treatment.
• 9:36
  Linas
• 9:37
  Beliūnas
• 9:37
  published
• 9:37
  a
• 9:37
  detailed
• 9:38
  framework
• 9:38
  for
• 9:38
  treating
• 9:39
  Anthropic's
• 9:39
  Cowork
• 9:40
  not
• 9:40
  as
• 9:40
  a
• 9:40
  chatbot
• 9:41
  with
• 9:41
  folder
• 9:41
  access
• 9:42
  but
• 9:42
  as
• 9:42
  a
• 9:42
  delegated
• 9:42
  operator.
• 9:44
  The
• 9:44
  key
• 9:44
  insight:
• 9:45
  stop
• 9:45
  writing
• 9:45
  step-by-step
• 9:46
  prompts
• 9:46
  and
• 9:46
  start
• 9:47
  describing
• 9:47
  outcomes.
• 9:49
  Cowork
• 9:49
  reads
• 9:49
  and
• 9:49
  writes
• 9:50
  files
• 9:50
  directly
• 9:50
  on
• 9:50
  your
• 9:51
  machine,
• 9:51
  spins
• 9:52
  up
• 9:52
  parallel
• 9:52
  workers
• 9:52
  for
• 9:53
  independent
• 9:53
  subtasks,
• 9:54
  and
• 9:54
  connects
• 9:54
  to
• 9:55
  50+
• 9:55
  tools
• 9:55
  via
• 9:56
  MCP
• 9:56
  (Slack,
• 9:57
  Google
• 9:57
  Drive,
• 9:57
  Notion,
• 9:58
  Jira,
• 9:58
  Snowflake,
• 9:59
  etc.).
• 10:00
  The
• 10:00
  framework
• 10:01
  emphasizes
• 10:01
  context
• 10:02
  files
• 10:02
  —
• 10:03
  three
• 10:03
  markdown
• 10:03
  documents
• 10:04
  describing
• 10:04
  who
• 10:04
  you
• 10:05
  are,
• 10:05
  how
• 10:05
  you
• 10:05
  communicate,
• 10:06
  and
• 10:06
  how
• 10:06
  you
• 10:07
  want
• 10:07
  Claude
• 10:07
  to
• 10:07
  behave
• 10:07
  —
• 10:08
  that
• 10:08
  compound
• 10:09
  over
• 10:09
  time.
• 10:10
  Beliūnas
• 10:11
  reports
• 10:11
  a
• 10:11
  user
• 10:12
  finding
• 10:12
  a
• 10:12
  $14,000/month
• 10:15
  pricing
• 10:15
  anomaly
• 10:16
  in
• 10:16
  45,000
• 10:17
  rows
• 10:17
  of
• 10:17
  revenue
• 10:17
  data
• 10:18
  using
• 10:18
  the
• 10:18
  Data
• 10:18
  Analyst
• 10:19
  plugin.
• 10:20
  The
• 10:20
  limitations
• 10:21
  are
• 10:21
  honest:
• 10:22
  no
• 10:22
  session
• 10:22
  persistence
• 10:23
  (every
• 10:23
  new
• 10:23
  session
• 10:24
  starts
• 10:24
  fresh),
• 10:25
  tasks
• 10:25
  die
• 10:26
  if
• 10:26
  you
• 10:26
  quit
• 10:26
  the
• 10:26
  app,
• 10:27
  and
• 10:27
  heavy
• 10:27
  multi-step
• 10:27
  tasks
• 10:28
  consume
• 10:29
  usage
• 10:29
  fast.
• 10:30
  Google
• 10:30
  Workspace
• 10:31
  now
• 10:31
  has
• 10:31
  a
• 10:31
  CLI
• 10:32
  called
• 10:32
  gws
• 10:32
  —
• 10:33
  not
• 10:33
  officially
• 10:34
  supported
• 10:34
  by
• 10:34
  Google,
• 10:35
  but
• 10:35
  living
• 10:35
  in
• 10:35
  the
• 10:35
  googleworkspace
• 10:36
  GitHub
• 10:36
  org.
• 10:37
  The
• 10:37
  clever
• 10:37
  design:
• 10:38
  gws
• 10:39
  doesn't
• 10:39
  ship
• 10:39
  a
• 10:39
  static
• 10:39
  list
• 10:40
  of
• 10:40
  commands.
• 10:41
  It
• 10:41
  reads
• 10:41
  Google's
• 10:42
  Discovery
• 10:42
  Service
• 10:42
  at
• 10:43
  runtime
• 10:43
  and
• 10:43
  builds
• 10:43
  its
• 10:44
  entire
• 10:44
  command
• 10:44
  surface
• 10:45
  dynamically,
• 10:45
  meaning
• 10:46
  it
• 10:46
  automatically
• 10:46
  picks
• 10:47
  up
• 10:47
  new
• 10:47
  Workspace
• 10:47
  API
• 10:48
  endpoints.
• 10:49
  It
• 10:49
  exposes
• 10:49
  100+
• 10:50
  agent
• 10:50
  skills
• 10:51
  via
• 10:51
  SKILL.md
• 10:52
  files,
• 10:53
  includes
• 10:53
  a
• 10:53
  built-in
• 10:53
  MCP
• 10:54
  server
• 10:54
  (gws
• 10:55
  mcp
• 10:55
  -s
• 10:56
  drive,gmail,calendar),
• 10:57
  and
• 10:57
  supports
• 10:57
  Model
• 10:58
  Armor
• 10:58
  response
• 10:58
  sanitization
• 10:59
  to
• 10:59
  scan
• 10:59
  API
• 10:59
  responses
• 11:00
  for
• 11:00
  prompt
• 11:00
  injection
• 11:01
  before
• 11:01
  they
• 11:01
  reach
• 11:01
  your
• 11:01
  agent.
• 11:02
  Authentication
• 11:03
  is
• 11:03
  encrypted
• 11:04
  at
• 11:04
  rest
• 11:04
  (AES-256-GCM)
• 11:07
  with
• 11:07
  keys
• 11:08
  in
• 11:08
  the
• 11:08
  OS
• 11:08
  keyring.
• 11:09
  For
• 11:09
  AI
• 11:09
  agents
• 11:10
  that
• 11:10
  need
• 11:10
  to
• 11:10
  manage
• 11:10
  Workspace,
• 11:11
  this
• 11:11
  eliminates
• 11:11
  substantial
• 11:12
  boilerplate.
• 11:13
  Rounding
• 11:13
  out
• 11:13
  the
• 11:14
  developer
• 11:14
  tooling:
• 11:15
  Nitpicker
• 11:15
  runs
• 11:15
  multiple
• 11:16
  LLM
• 11:16
  reviewers
• 11:17
  concurrently
• 11:17
  on
• 11:17
  a
• 11:17
  git
• 11:18
  repository,
• 11:19
  then
• 11:19
  aggregates
• 11:19
  their
• 11:19
  findings
• 11:20
  into
• 11:20
  a
• 11:20
  single
• 11:21
  prioritized
• 11:21
  report.
• 11:22
  Each
• 11:22
  reviewer
• 11:23
  is
• 11:23
  an
• 11:23
  agentic
• 11:23
  loop
• 11:23
  with
• 11:24
  tool
• 11:24
  access
• 11:24
  (read
• 11:25
  files,
• 11:25
  grep,
• 11:25
  glob,
• 11:26
  git
• 11:26
  commands),
• 11:27
  and
• 11:27
  a
• 11:27
  separate
• 11:27
  aggregator
• 11:28
  model
• 11:28
  deduplicates
• 11:29
  and
• 11:29
  synthesizes
• 11:29
  the
• 11:29
  reviews.
• 11:30
  The
• 11:31
  recommended
• 11:31
  configuration
• 11:32
  uses
• 11:32
  different
• 11:32
  providers
• 11:33
  to
• 11:33
  enforce
• 11:33
  diversity
• 11:34
  of
• 11:34
  thought
• 11:34
  —
• 11:35
  Claude
• 11:35
  for
• 11:35
  one
• 11:36
  reviewer,
• 11:36
  GPT
• 11:36
  for
• 11:37
  another,
• 11:37
  Gemini
• 11:38
  as
• 11:38
  aggregator.
• 11:39
  It's
• 11:39
  a
• 11:39
  Rust
• 11:39
  CLI,
• 11:40
  minimal
• 11:40
  and
• 11:40
  focused.
• 11:42
  Meanwhile,
• 11:42
  a
• 11:42
  father-son
• 11:43
  team
• 11:43
  documented
• 11:44
  their
• 11:44
  experience
• 11:44
  ramping
• 11:45
  up
• 11:45
  on
• 11:45
  Gas
• 11:45
  Town,
• 11:46
  an
• 11:46
  agentic
• 11:47
  AI
• 11:47
  orchestrator.
• 11:48
  Their
• 11:48
  key
• 11:49
  discovery:
• 11:49
  don't
• 11:50
  issue
• 11:50
  CLI
• 11:51
  commands
• 11:51
  yourself
• 11:51
  —
• 11:52
  ask
• 11:53
  the
• 11:53
  "Mayor"
• 11:53
  agent
• 11:53
  in
• 11:54
  conversational
• 11:54
  English
• 11:55
  and
• 11:55
  let
• 11:55
  it
• 11:55
  delegate
• 11:56
  to
• 11:56
  "Polecat"
• 11:56
  coding
• 11:56
  agents.
• 11:58
  The
• 11:58
  pattern
• 11:58
  mirrors
• 11:58
  a
• 11:58
  broader
• 11:59
  shift:
• 12:00
  the
• 12:00
  most
• 12:00
  effective
• 12:00
  way
• 12:01
  to
• 12:01
  use
• 12:01
  multi-agent
• 12:01
  systems
• 12:02
  is
• 12:02
  to
• 12:02
  stay
• 12:03
  at
• 12:03
  the
• 12:03
  intent
• 12:03
  level
• 12:04
  and
• 12:04
  let
• 12:04
  the
• 12:04
  orchestration
• 12:04
  layer
• 12:05
  handle
• 12:05
  execution.
• 12:10
  Dhillon
• 12:10
  Andrew
• 12:10
  Kannabhiran
• 12:11
  published
• 12:11
  (S)AGE
• 12:11
  —
• 12:12
  Sovereign
• 12:12
  Agent
• 12:13
  Governed
• 12:13
  Experience
• 12:13
  —
• 12:14
  an
• 12:14
  open-source
• 12:14
  infrastructure
• 12:15
  layer
• 12:15
  for
• 12:15
  governed,
• 12:16
  verifiable,
• 12:17
  experience-weighted
• 12:17
  institutional
• 12:18
  memory
• 12:18
  across
• 12:18
  multi-agent
• 12:19
  systems.
• 12:20
  The
• 12:20
  motivation:
• 12:21
  existing
• 12:21
  agent
• 12:22
  memory
• 12:22
  systems
• 12:22
  (Mem0,
• 12:23
  MemGPT,
• 12:24
  Reflexion)
• 12:24
  all
• 12:25
  measure
• 12:25
  retrieval
• 12:25
  accuracy
• 12:26
  and
• 12:26
  latency
• 12:26
  but
• 12:26
  never
• 12:26
  measure
• 12:27
  whether
• 12:27
  memory
• 12:27
  actually
• 12:27
  makes
• 12:28
  agents
• 12:28
  produce
• 12:28
  better
• 12:28
  outcomes.
• 12:29
  And
• 12:29
  they
• 12:29
  all
• 12:30
  do
• 12:30
  the
• 12:30
  same
• 12:30
  thing:
• 12:30
  store
• 12:31
  everything,
• 12:31
  retrieve
• 12:32
  what
• 12:32
  seems
• 12:32
  relevant.
• 12:33
  Memory
• 12:33
  without
• 12:33
  judgment.
• 12:34
  (S)AGE
• 12:34
  implements
• 12:34
  a
• 12:34
  governance
• 12:35
  pipeline:
• 12:36
  agents
• 12:36
  propose
• 12:36
  observations,
• 12:37
  a
• 12:37
  4-node
• 12:38
  BFT
• 12:38
  consensus
• 12:39
  network
• 12:39
  validates
• 12:39
  them
• 12:39
  using
• 12:40
  Proof
• 12:40
  of
• 12:40
  Experience
• 12:40
  (reputation-weighted
• 12:41
  voting
• 12:41
  with
• 12:42
  domain
• 12:42
  relevance,
• 12:42
  accuracy,
• 12:43
  recency,
• 12:44
  and
• 12:44
  corroboration
• 12:45
  factors),
• 12:45
  and
• 12:46
  committed
• 12:46
  knowledge
• 12:47
  becomes
• 12:47
  institutional
• 12:47
  truth
• 12:48
  that
• 12:48
  can
• 12:48
  be
• 12:48
  challenged
• 12:49
  with
• 12:49
  evidence
• 12:49
  later.
• 12:50
  The
• 12:50
  stack
• 12:51
  runs
• 12:51
  entirely
• 12:51
  local
• 12:52
  —
• 12:52
  CometBFT
• 12:53
  for
• 12:53
  consensus,
• 12:53
  PostgreSQL
• 12:54
  with
• 12:54
  pgvector
• 12:55
  for
• 12:55
  storage,
• 12:56
  Ollama
• 12:56
  for
• 12:56
  embeddings,
• 12:57
  no
• 12:57
  cloud
• 12:57
  calls.
• 12:58
  Performance:
• 12:58
  956
• 13:00
  req/s
• 13:01
  submissions,
• 13:02
  21.6ms
• 13:03
  P95
• 13:03
  queries.
• 13:04
  The
• 13:04
  threat
• 13:05
  model
• 13:05
  is
• 13:05
  thorough:
• 13:05
  memory
• 13:06
  poisoning,
• 13:07
  Sybil
• 13:07
  attacks,
• 13:07
  collusion
• 13:08
  detection
• 13:08
  via
• 13:08
  pairwise
• 13:09
  voting
• 13:09
  correlation,
• 13:10
  prompt
• 13:10
  injection
• 13:10
  via
• 13:11
  memory
• 13:11
  (content
• 13:11
  rendered
• 13:12
  as
• 13:12
  data,
• 13:12
  never
• 13:12
  instructions).
• 13:13
  Whether
• 13:14
  organizations
• 13:14
  will
• 13:15
  deploy
• 13:15
  BFT
• 13:15
  consensus
• 13:16
  for
• 13:16
  agent
• 13:16
  knowledge
• 13:17
  remains
• 13:17
  an
• 13:17
  open
• 13:17
  question,
• 13:18
  but
• 13:18
  the
• 13:18
  framework
• 13:19
  identifies
• 13:19
  a
• 13:19
  real
• 13:19
  gap:
• 13:20
  memory
• 13:20
  governance
• 13:21
  for
• 13:21
  multi-agent
• 13:21
  systems.
• 13:23
  Open
• 13:23
  Trajectory
• 13:23
  Gym
• 13:24
  tackles
• 13:24
  a
• 13:24
  different
• 13:24
  gap:
• 13:25
  training
• 13:25
  agents
• 13:25
  to
• 13:26
  execute,
• 13:26
  not
• 13:26
  just
• 13:27
  converse.
• 13:28
  The
• 13:28
  framework
• 13:29
  treats
• 13:29
  entire
• 13:29
  agent
• 13:30
  traces
• 13:30
  —
• 13:31
  observation,
• 13:31
  tool
• 13:32
  call,
• 13:32
  outcome
• 13:32
  sequences
• 13:33
  —
• 13:34
  as
• 13:34
  first-class
• 13:34
  training
• 13:35
  data,
• 13:35
  using
• 13:36
  three
• 13:36
  complementary
• 13:36
  methods.
• 13:38
  Supervised
• 13:38
  fine-tuning
• 13:39
  on
• 13:39
  successful
• 13:39
  traces
• 13:40
  improves
• 13:40
  CyBench
• 13:41
  CTF
• 13:41
  solve
• 13:41
  rates
• 13:41
  from
• 13:42
  12.5%
• 13:42
  to
• 13:43
  roughly
• 13:43
  20%.
• 13:44
  Online
• 13:45
  RL
• 13:45
  with
• 13:45
  live
• 13:45
  tool
• 13:45
  execution
• 13:46
  (using
• 13:46
  a
• 13:46
  patched
• 13:47
  SkyRL
• 13:47
  fork)
• 13:47
  pushes
• 13:48
  to
• 13:48
  ~29%.
• 13:49
  But
• 13:50
  the
• 13:50
  surprise
• 13:50
  result
• 13:50
  is
• 13:51
  GEPA
• 13:51
  —
• 13:52
  Genetic
• 13:52
  Prompt
• 13:52
  Evolution
• 13:53
  with
• 13:53
  Pareto
• 13:53
  optimization
• 13:54
  —
• 13:54
  which
• 13:55
  achieves
• 13:55
  ~35%
• 13:56
  solve
• 13:56
  rates
• 13:57
  using
• 13:57
  4–35x
• 13:59
  fewer
• 13:59
  compute
• 13:59
  resources
• 14:00
  than
• 14:00
  RL,
• 14:00
  evolving
• 14:01
  prompts
• 14:01
  rather
• 14:01
  than
• 14:02
  updating
• 14:02
  weights.
• 14:03
  The
• 14:03
  caveat:
• 14:03
  GEPA
• 14:04
  only
• 14:04
  surfaces
• 14:04
  capability
• 14:05
  the
• 14:05
  model
• 14:05
  already
• 14:06
  has.
• 14:06
  The
• 14:06
  resource
• 14:07
  requirements
• 14:07
  are
• 14:07
  substantial
• 14:08
  (140GB+
• 14:10
  VRAM
• 14:10
  for
• 14:10
  Online
• 14:11
  RL),
• 14:11
  and
• 14:11
  the
• 14:12
  framework
• 14:12
  is
• 14:12
  at
• 14:12
  v0.1.0
• 14:14
  with
• 14:14
  a
• 14:14
  single
• 14:14
  benchmark
• 14:15
  (CyBench
• 14:15
  CTF).
• 14:16
  The
• 14:16
  broader
• 14:17
  question
• 14:17
  it
• 14:17
  raises
• 14:17
  is
• 14:17
  whether
• 14:18
  prompt
• 14:18
  engineering
• 14:18
  vs.
• 14:19
  gradient
• 14:19
  optimization
• 14:20
  will
• 14:20
  remain
• 14:20
  a
• 14:20
  useful
• 14:20
  distinction
• 14:21
  as
• 14:21
  agent
• 14:21
  post-training
• 14:22
  matures.
• 14:23
  On
• 14:23
  the
• 14:23
  inference
• 14:24
  side,
• 14:24
  a
• 14:24
  new
• 14:25
  arXiv
• 14:25
  paper
• 14:25
  introduces
• 14:26
  Speculative
• 14:26
  Speculative
• 14:27
  Decoding
• 14:27
  (SSD)
• 14:28
  —
• 14:29
  parallelizing
• 14:29
  even
• 14:30
  the
• 14:30
  sequential
• 14:30
  dependence
• 14:30
  between
• 14:31
  drafting
• 14:31
  and
• 14:31
  verification
• 14:32
  in
• 14:32
  standard
• 14:33
  speculative
• 14:33
  decoding.
• 14:34
  While
• 14:34
  verification
• 14:35
  runs
• 14:35
  on
• 14:35
  the
• 14:35
  target
• 14:36
  model,
• 14:36
  the
• 14:36
  draft
• 14:37
  model
• 14:37
  predicts
• 14:37
  likely
• 14:38
  verification
• 14:38
  outcomes
• 14:39
  and
• 14:39
  pre-computes
• 14:39
  speculations
• 14:40
  for
• 14:40
  each.
• 14:41
  If
• 14:41
  the
• 14:41
  actual
• 14:42
  outcome
• 14:42
  matches
• 14:42
  a
• 14:42
  cached
• 14:43
  prediction,
• 14:43
  the
• 14:44
  speculation
• 14:44
  returns
• 14:45
  immediately
• 14:45
  with
• 14:45
  zero
• 14:46
  draft
• 14:46
  overhead.
• 14:47
  The
• 14:47
  optimized
• 14:47
  algorithm,
• 14:48
  Saguaro,
• 14:49
  uses
• 14:49
  geometric
• 14:50
  fan-out
• 14:50
  allocation,
• 14:51
  a
• 14:51
  novel
• 14:51
  cache-aware
• 14:52
  sampling
• 14:52
  scheme
• 14:53
  that
• 14:53
  trades
• 14:53
  acceptance
• 14:53
  rate
• 14:54
  for
• 14:54
  cache
• 14:54
  hit
• 14:54
  rate,
• 14:55
  and
• 14:55
  an
• 14:55
  adaptive
• 14:55
  backup
• 14:56
  strategy
• 14:56
  that
• 14:56
  switches
• 14:57
  speculators
• 14:57
  based
• 14:58
  on
• 14:58
  batch
• 14:58
  size.
• 14:59
  Results:
• 15:00
  up
• 15:00
  to
• 15:00
  2x
• 15:01
  faster
• 15:01
  than
• 15:01
  optimized
• 15:02
  speculative
• 15:02
  decoding,
• 15:03
  up
• 15:03
  to
• 15:03
  5x
• 15:04
  over
• 15:04
  autoregressive
• 15:05
  generation,
• 15:06
  while
• 15:06
  improving
• 15:07
  the
• 15:07
  throughput-latency
• 15:07
  Pareto
• 15:08
  frontier.
• 15:09
  The
• 15:09
  key
• 15:09
  constraint:
• 15:10
  the
• 15:10
  draft
• 15:11
  model
• 15:11
  must
• 15:11
  run
• 15:11
  on
• 15:12
  separate
• 15:12
  hardware
• 15:12
  from
• 15:12
  the
• 15:12
  target,
• 15:13
  making
• 15:13
  this
• 15:14
  primarily
• 15:14
  a
• 15:14
  data-center
• 15:15
  optimization.
• 15:15
  The
• 15:16
  Wave-Field
• 15:16
  LLM,
• 15:17
  a
• 15:17
  GitHub
• 15:17
  project
• 15:18
  proposing
• 15:18
  O(n
• 15:19
  log
• 15:19
  n)
• 15:19
  language
• 15:19
  modeling
• 15:20
  via
• 15:20
  wave
• 15:20
  equation
• 15:21
  dynamics,
• 15:22
  offers
• 15:22
  a
• 15:22
  more
• 15:22
  speculative
• 15:23
  architectural
• 15:23
  departure,
• 15:24
  though
• 15:24
  details
• 15:25
  remain
• 15:25
  sparse.
• 15:28
  DeepNote
• 15:29
  AI
• 15:29
  dropped
• 15:29
  as
• 15:30
  an
• 15:30
  open-source
• 15:30
  desktop
• 15:31
  application
• 15:31
  inspired
• 15:32
  by
• 15:32
  Google's
• 15:32
  NotebookLM,
• 15:33
  built
• 15:33
  with
• 15:34
  Electron,
• 15:34
  React,
• 15:35
  and
• 15:35
  multi-provider
• 15:36
  AI
• 15:36
  (Gemini,
• 15:37
  Claude,
• 15:37
  OpenAI,
• 15:38
  Groq).
• 15:39
  The
• 15:40
  feature
• 15:40
  list
• 15:40
  is
• 15:40
  ambitious
• 15:41
  for
• 15:41
  a
• 15:41
  beta:
• 15:43
  upload
• 15:43
  documents
• 15:43
  and
• 15:44
  chat
• 15:44
  with
• 15:44
  sources
• 15:44
  via
• 15:45
  agentic
• 15:45
  RAG
• 15:46
  (multi-query
• 15:46
  retrieval
• 15:47
  with
• 15:47
  sub-query
• 15:48
  generation
• 15:49
  and
• 15:49
  sufficiency
• 15:49
  checking),
• 15:50
  generate
• 15:51
  AI
• 15:51
  podcasts
• 15:52
  with
• 15:52
  multi-speaker
• 15:53
  TTS,
• 15:54
  build
• 15:55
  image
• 15:55
  slide
• 15:55
  decks
• 15:56
  with
• 15:56
  a
• 15:56
  drag-and-drop
• 15:56
  editor,
• 15:57
  create
• 15:57
  whitepapers,
• 15:58
  infographics,
• 15:59
  flashcards,
• 16:00
  quizzes,
• 16:01
  mind
• 16:01
  maps,
• 16:02
  dashboards,
• 16:03
  literature
• 16:04
  reviews,
• 16:04
  and
• 16:05
  competitive
• 16:05
  analyses.
• 16:06
  Complex
• 16:07
  content
• 16:07
  types
• 16:07
  use
• 16:08
  a
• 16:08
  Research
• 16:08
  →
• 16:08
  Write
• 16:08
  →
• 16:09
  Review
• 16:10
  multi-agent
• 16:10
  pipeline
• 16:11
  with
• 16:11
  automatic
• 16:11
  revision
• 16:12
  if
• 16:12
  quality
• 16:12
  scores
• 16:12
  below
• 16:13
  6/10.
• 16:14
  Cross-session
• 16:14
  memory
• 16:14
  persists
• 16:15
  user
• 16:15
  preferences
• 16:16
  per
• 16:16
  notebook.
• 16:17
  Local
• 16:17
  ONNX
• 16:17
  embeddings
• 16:18
  (all-MiniLM-L6-v2)
• 16:19
  enable
• 16:19
  offline
• 16:20
  RAG
• 16:20
  with
• 16:20
  tiered
• 16:20
  fallback
• 16:21
  to
• 16:21
  Gemini
• 16:21
  API
• 16:22
  to
• 16:22
  hash-based
• 16:22
  embeddings.
• 16:23
  It
• 16:23
  even
• 16:23
  includes
• 16:24
  a
• 16:24
  DeepBrain
• 16:24
  integration
• 16:25
  for
• 16:25
  system-wide
• 16:25
  memory
• 16:25
  recall
• 16:26
  and
• 16:26
  file
• 16:26
  search.
• 16:27
  The
• 16:28
  known
• 16:28
  issues
• 16:28
  are
• 16:28
  typical
• 16:29
  beta:
• 16:30
  linear
• 16:30
  vector
• 16:30
  search,
• 16:31
  unbounded
• 16:31
  cache
• 16:32
  growth,
• 16:32
  API
• 16:33
  keys
• 16:33
  stored
• 16:33
  as
• 16:33
  plaintext.
• 16:35
  But
• 16:35
  as
• 16:35
  a
• 16:35
  demonstration
• 16:35
  of
• 16:36
  how
• 16:36
  much
• 16:36
  AI-native
• 16:37
  desktop
• 16:37
  tooling
• 16:38
  can
• 16:38
  pack
• 16:38
  into
• 16:38
  a
• 16:38
  single
• 16:38
  application,
• 16:39
  it's
• 16:39
  worth
• 16:39
  watching.
• 16:41
  The
• 16:41
  generalist-specialist
• 16:42
  debate
• 16:42
  is
• 16:42
  getting
• 16:42
  a
• 16:42
  rewrite
• 16:43
  from
• 16:43
  AI.
• 16:44
  Rick
• 16:44
  Deacon
• 16:44
  argues
• 16:45
  that
• 16:45
  agentic
• 16:45
  AI
• 16:46
  compresses
• 16:46
  the
• 16:46
  gap
• 16:46
  between
• 16:47
  vision
• 16:47
  and
• 16:47
  execution,
• 16:48
  turning
• 16:49
  generalists
• 16:49
  from
• 16:49
  synthesizers
• 16:50
  into
• 16:50
  builders
• 16:50
  with
• 16:50
  direct
• 16:51
  execution
• 16:51
  leverage.
• 16:52
  The
• 16:53
  pattern
• 16:53
  is
• 16:53
  familiar
• 16:53
  to
• 16:54
  anyone
• 16:54
  who's
• 16:54
  worked
• 16:54
  across
• 16:54
  offensive
• 16:55
  security,
• 16:55
  application
• 16:56
  security,
• 16:57
  product,
• 16:57
  and
• 16:57
  startup
• 16:57
  building:
• 16:58
  breadth
• 16:58
  lets
• 16:58
  you
• 16:59
  connect
• 16:59
  dots,
• 16:59
  lead
• 17:00
  teams,
• 17:00
  and
• 17:00
  translate
• 17:01
  between
• 17:01
  engineers
• 17:01
  and
• 17:01
  executives,
• 17:02
  but
• 17:02
  historically
• 17:03
  you
• 17:03
  still
• 17:03
  needed
• 17:03
  specialists
• 17:04
  to
• 17:04
  execute.
• 17:05
  With
• 17:05
  LLMs
• 17:06
  and
• 17:06
  autonomous
• 17:06
  agents,
• 17:07
  generalists
• 17:08
  can
• 17:08
  prototype
• 17:08
  ideas,
• 17:09
  spin
• 17:09
  up
• 17:10
  frameworks,
• 17:10
  simulate
• 17:11
  security
• 17:11
  models,
• 17:12
  and
• 17:12
  iterate
• 17:12
  —
• 17:13
  all
• 17:13
  directly.
• 17:14
  The
• 17:14
  risk
• 17:14
  is
• 17:14
  overconfidence:
• 17:15
  being
• 17:16
  able
• 17:16
  to
• 17:16
  prototype
• 17:16
  isn't
• 17:17
  the
• 17:17
  same
• 17:17
  as
• 17:17
  being
• 17:17
  able
• 17:18
  to
• 17:18
  ship
• 17:18
  production-quality
• 17:19
  work,
• 17:19
  and
• 17:19
  the
• 17:19
  "multi-tool
• 17:20
  with
• 17:20
  every
• 17:20
  edge
• 17:21
  sharp"
• 17:21
  metaphor
• 17:21
  glosses
• 17:22
  over
• 17:22
  the
• 17:22
  domains
• 17:22
  where
• 17:23
  deep
• 17:23
  specialization
• 17:23
  still
• 17:24
  dominates.
• 17:25
  But
• 17:25
  the
• 17:25
  directional
• 17:26
  claim
• 17:26
  is
• 17:26
  sound:
• 17:27
  when
• 17:27
  tools
• 17:27
  lower
• 17:28
  the
• 17:28
  cost
• 17:28
  of
• 17:28
  specialization,
• 17:29
  the
• 17:29
  advantage
• 17:30
  shifts
• 17:30
  to
• 17:30
  those
• 17:30
  who
• 17:31
  understand
• 17:31
  the
• 17:31
  whole
• 17:31
  system.
• 17:32
  OCR-Provenance,
• 17:33
  a
• 17:34
  GitHub
• 17:34
  project
• 17:34
  for
• 17:34
  document
• 17:35
  provenance
• 17:35
  verification
• 17:36
  using
• 17:36
  OCR,
• 17:37
  aims
• 17:37
  to
• 17:38
  address
• 17:38
  the
• 17:38
  authenticity
• 17:39
  gap
• 17:39
  in
• 17:39
  an
• 17:39
  era
• 17:39
  of
• 17:39
  AI-generated
• 17:40
  documents,
• 17:41
  though
• 17:41
  the
• 17:42
  repository
• 17:42
  remains
• 17:43
  sparse
• 17:43
  on
• 17:43
  implementation
• 17:44
  details.