WEBVTT

1
00:00:00.000 --> 00:00:04.807
[Nina] A supply chain worm hid inside legitimate packages for twenty-six days.

2
00:00:04.807 --> 00:00:13.549
This week it came back, and the commits it pushes to your repos are authored as an AI coding assistant.

3
00:00:13.549 --> 00:00:14.861
That's Mini Shai-Hulud.

4
00:00:31.780 --> 00:00:38.360
[Max] Welcome to Context Window. Your weekly window into everything worth knowing in AI security.

5
00:00:38.360 --> 00:00:43.530
Quick thing before we start. Full transparency: this podcast is AI-generated.

6
00:00:43.530 --> 00:00:49.640
Two AI voices, a script written by another AI agent, the whole thing.

7
00:00:49.640 --> 00:00:56.220
But every story, every take, every insight is curated and verified by Assaf Nakash.

8
00:00:56.220 --> 00:01:02.330
He builds AI security products at one of the world's largest security platforms.

9
00:01:02.330 --> 00:01:10.790
Yes, the robots do the talking, but a human who lives this stuff makes sure we're not hallucinating.

10
00:01:11.040 --> 00:01:16.630
[Nina] If we ever do start hallucinating, the worm will probably commit it under our name anyway.

11
00:01:16.630 --> 00:01:18.376
Until then, let Assaf know.

12
00:01:18.626 --> 00:01:20.227
[Max] Three things this week.

13
00:01:20.227 --> 00:01:29.033
A supply chain worm that hit two package ecosystems at the same time and signed its commits as an AI coding assistant.

14
00:01:29.033 --> 00:01:35.036
A Linux kernel vulnerability that corrupts memory while your file integrity monitors report all clear.

15
00:01:35.036 --> 00:01:40.640
And Assaf's thesis on why trust itself has become the exploit primitive of 2026.

16
00:01:40.890 --> 00:01:43.954
[Nina] That last one ties everything together. Let's get into it.

17
00:01:47.355 --> 00:01:55.204
[Max] Top story. TeamPCP. They're a threat actor group running coordinated supply chain attacks through open-source package registries.

18
00:01:55.204 --> 00:02:01.668
Back in March, they launched what researchers called Shai-Hulud. Quick context for non-Dune fans.

19
00:02:01.668 --> 00:02:08.132
Named after the giant sandworm. A self-propagating worm that spread through compromised npm packages.

20
00:02:08.132 --> 00:02:11.364
Then they went quiet for twenty-six days.

21
00:02:11.614 --> 00:02:14.122
[Nina] And now they're back with a sequel.

22
00:02:14.372 --> 00:02:23.319
[Max] April twenty-ninth, compromised SAP npm packages started appearing. cap-js db-service, cap-js postgres. Real packages maintained by SAP.

23
00:02:23.319 --> 00:02:30.687
Malicious preinstall scripts harvesting credentials from the build environment. One day later, it escalated.

24
00:02:30.687 --> 00:02:38.582
PyTorch Lightning on PyPI was compromised. Intercom's npm client was hit at the same time.

25
00:02:38.582 --> 00:02:47.530
Researchers at Kodem Security are calling this variant Mini Shai-Hulud because it attacks across two ecosystems simultaneously.

26
00:02:47.780 --> 00:02:54.884
[Nina] Both legitimate packages from legitimate publisher accounts. Not typosquats. Not lookalike names.

27
00:02:55.134 --> 00:03:00.355
[Max] The Python package executes on import. The npm package executes on install.

28
00:03:00.355 --> 00:03:07.316
Both download an obfuscated payload via the Bun runtime and harvest every credential in the environment.

29
00:03:07.316 --> 00:03:11.667
GitHub, npm, cloud providers, Kubernetes, CI/CD variables. All of it.

30
00:03:11.917 --> 00:03:13.170
[Nina] That's a wide net.

31
00:03:13.420 --> 00:03:15.229
[Max] Then the worm propagates.

32
00:03:15.229 --> 00:03:22.011
For every stolen token, it commits a copy of itself to up to fifty branches.

33
00:03:22.011 --> 00:03:28.793
And those commits are authored as claude at users dot noreply dot github dot com.

34
00:03:29.043 --> 00:03:33.919
[Nina] So the malicious commits look like they came from an AI coding assistant.

35
00:03:34.169 --> 00:03:41.405
[Max] That's the social engineering layer. It hides in plain sight because AI-authored commits are normal now.

36
00:03:41.405 --> 00:03:46.832
PyPI has quarantined the malicious lightning package. npm pulled intercom-client version 7.0.4.

37
00:03:46.832 --> 00:03:54.973
But any token stolen during the exposure window is still live. The worm already propagated with those tokens.

38
00:03:55.223 --> 00:03:58.752
[Nina] Removing the package doesn't recall the credentials.

39
00:03:59.002 --> 00:04:07.990
[Max] Unit 42 published attribution connecting this back to TeamPCP's earlier campaigns. Same infrastructure. Same harvesting techniques.

40
00:04:07.990 --> 00:04:10.799
This is organized and sustained.

41
00:04:11.048 --> 00:04:13.417
[Nina] What should teams be doing right now?

42
00:04:13.667 --> 00:04:17.644
[Max] Check whether your environments installed any of the affected packages.

43
00:04:17.644 --> 00:04:21.223
Rotate every credential that was accessible to those environments.

44
00:04:21.223 --> 00:04:25.598
And search your git history for commits from that author address.

45
00:04:25.598 --> 00:04:29.177
If you find them, assume the branch is compromised.

46
00:04:29.427 --> 00:04:37.694
[Nina] These packages passed every check. Real publishers. Legitimate accounts. The green checkmark was genuine.

47
00:04:37.944 --> 00:04:39.290
[Max] And it was the weapon.

48
00:04:42.689 --> 00:04:45.197
[Max] Okay, speed round. What else happened this week.

49
00:04:45.447 --> 00:04:53.518
[Max] First up. A Linux kernel vulnerability called Copy Fail. A logic flaw in the kernel's cryptographic interface.

50
00:04:53.518 --> 00:04:56.841
It's been in the kernel since 2017.

51
00:04:56.841 --> 00:05:06.810
Any unprivileged local user can corrupt the page cache of system binaries that run with root privileges to gain root access.

52
00:05:07.060 --> 00:05:11.704
[Nina] Any unprivileged user. How bad is the proof-of-concept?

53
00:05:11.954 --> 00:05:17.003
[Max] Seven hundred and thirty-two bytes of Python. It crosses container boundaries.

54
00:05:17.003 --> 00:05:24.808
But here's the part that should keep you up at night. The corruption happens entirely in memory.

55
00:05:24.808 --> 00:05:28.021
The actual files on disk stay untouched.

56
00:05:28.271 --> 00:05:32.220
[Nina] Your integrity tooling says all clear while someone already has root.

57
00:05:32.470 --> 00:05:37.550
[Max] The trust in the verification mechanism is the blind spot. Upstream patches have landed.

58
00:05:37.550 --> 00:05:40.452
Most Linux distributions are still rolling the fixes.

59
00:05:40.452 --> 00:05:44.807
In the meantime, you can blacklist the affected kernel module to mitigate.

60
00:05:44.807 --> 00:05:48.072
The specific module name is in the show notes.

61
00:05:48.322 --> 00:05:49.804
[Max] Next story. LiteLLM.

62
00:05:49.804 --> 00:05:58.202
This is the same AI gateway that was compromised by TeamPCP's supply chain worm back in March.

63
00:05:58.202 --> 00:06:04.623
Different vulnerability class this time. A pre-auth SQL injection in the LiteLLM Proxy.

64
00:06:04.874 --> 00:06:07.939
[Nina] Pre-auth. No credentials needed to exploit it.

65
00:06:08.187 --> 00:06:17.661
[Max] The proxy takes the Bearer token from the authorization header and drops it straight into a SQL query.

66
00:06:17.661 --> 00:06:27.135
No sanitization. An unauthenticated attacker can reach the stored upstream API keys. OpenAI keys. Anthropic keys. AWS credentials.

67
00:06:27.135 --> 00:06:30.293
Everything LiteLLM is configured to proxy.

68
00:06:30.543 --> 00:06:36.209
[Nina] Sysdig says they saw targeted exploitation thirty-six hours after disclosure.

69
00:06:36.459 --> 00:06:44.906
[Max] Thirty-six hours. Patched in version 1.83.7. But think about the pattern. Same infrastructure, hit twice in six weeks.

70
00:06:44.906 --> 00:06:48.660
In March, the supply chain worm compromised it.

71
00:06:48.660 --> 00:06:53.823
In May, a completely different vulnerability class exposes the same keys.

72
00:06:53.823 --> 00:06:58.046
LiteLLM is what teams use to centralize model access.

73
00:06:58.046 --> 00:07:03.209
That centralization of trust is what makes it a high-value target.

74
00:07:03.459 --> 00:07:07.916
[Nina] The more you trust it with, the more valuable it becomes to an attacker.

75
00:07:08.166 --> 00:07:12.103
[Max] Last item. The AI security industry is consolidating fast.

76
00:07:12.103 --> 00:07:20.415
Palo Alto Networks announced it's acquiring Portkey for a reported one hundred twenty to one hundred forty million dollars.

77
00:07:20.415 --> 00:07:26.101
Portkey is an AI gateway that already processes trillions of tokens per month.

78
00:07:26.101 --> 00:07:29.601
It becomes the control plane for Prisma AIRS.

79
00:07:29.601 --> 00:07:35.288
This follows their acquisition of Koi earlier this year, the agentic endpoints startup.

80
00:07:35.536 --> 00:07:40.413
[Nina] They're assembling the entire AI security stack through M and A.

81
00:07:40.663 --> 00:07:47.968
[Max] Meanwhile, OpenAI launched GPT-5.5-Cyber and immediately locked it behind their Trusted Access for Cyber program.

82
00:07:47.968 --> 00:07:56.246
Only vetted government defenders, critical infrastructure operators, and select security vendors can use it. No public API.

83
00:07:56.246 --> 00:07:57.706
No published benchmarks.

84
00:07:57.956 --> 00:08:06.269
[Nina] Interesting, because OpenAI previously questioned Anthropic for restricting Mythos. Now they're taking the same approach.

85
00:08:06.519 --> 00:08:12.387
[Max] And the frontier labs are also learning to cooperate on defense.

86
00:08:12.387 --> 00:08:19.323
OpenAI, Anthropic, and Google formed a threat-intel sharing agreement targeting unauthorized model extraction.

87
00:08:19.323 --> 00:08:28.392
The trigger: Anthropic reported twenty-four thousand fraudulent accounts that extracted over sixteen million training exchanges from Claude.

88
00:08:28.642 --> 00:08:35.975
[Nina] Congress responded with the Stop AI Model Theft Act. Restrict access and share threat intel.

89
00:08:35.975 --> 00:08:37.931
That's the emerging playbook.

90
00:08:38.179 --> 00:08:41.275
[Max] Whether it works is an open question.

91
00:08:41.275 --> 00:08:47.467
But responsible release norms for cyber-capable models are now the standard across frontier labs.

92
00:08:50.067 --> 00:08:57.967
[Max] Alright, time for Curator's Corner. The if-you-listen-to-nothing-else moment. Assaf's thesis this week is straightforward.

93
00:08:57.967 --> 00:09:01.352
Trust has become an exploit primitive.

94
00:09:01.602 --> 00:09:03.600
[Nina] Bold claim. Walk us through it.

95
00:09:03.850 --> 00:09:12.936
[Max] The argument starts with this week's top story. TeamPCP's Mini Shai-Hulud compromised real packages from real publisher accounts.

96
00:09:12.936 --> 00:09:18.488
The artifact looked legitimate because it was legitimate. Until it wasn't.

97
00:09:18.488 --> 00:09:27.070
Organizations that auto-update from trusted sources were exposed first. Precisely because their trust was the most complete.

98
00:09:27.320 --> 00:09:32.381
[Nina] The fastest adopters of best practices got hit first. That's a rough irony.

99
00:09:32.630 --> 00:09:37.970
[Max] Copy Fail corrupts the page cache in memory while leaving the disk untouched.

100
00:09:37.970 --> 00:09:40.434
Your file integrity monitor reports clean.

101
00:09:40.434 --> 00:09:45.774
The more you trust your integrity tooling, the more invisible this attack becomes.

102
00:09:46.024 --> 00:09:49.134
[Nina] The verification mechanism is the blind spot.

103
00:09:49.384 --> 00:09:55.663
[Max] And LiteLLM has been hit twice in six weeks. Different vulnerability classes, same outcome.

104
00:09:55.663 --> 00:10:00.597
Credential exposure from the AI infrastructure teams trust to centralize access.

105
00:10:00.597 --> 00:10:05.081
The centralization of trust is what makes it the target.

106
00:10:05.331 --> 00:10:11.136
[Nina] Three stories. Three different trust assumptions. All exploited.

107
00:10:11.386 --> 00:10:18.589
[Max] The argument goes further. The way he sees it, three old security reflexes have broken.

108
00:10:18.589 --> 00:10:24.350
Updating to latest became a weapon because TeamPCP exploits the update path.

109
00:10:24.350 --> 00:10:31.072
Credential rotation can't outrun an automated worm that propagates to fifty branches per token.

110
00:10:31.072 --> 00:10:38.274
And security tools became the attack surface because LiteLLM's authentication path became the exfiltration path.

111
00:10:38.524 --> 00:10:42.565
[Nina] Those sound like predictions. But this week, they're the news.

112
00:10:42.815 --> 00:10:51.031
[Max] The reactive defenses still work. PyPI quarantined the malicious packages within hours. Copy Fail has module-blacklisting workarounds.

113
00:10:51.031 --> 00:10:57.798
Those are worth acknowledging. But the detection is reactive. The trust exploitation is structural.

114
00:10:57.798 --> 00:11:08.914
We catch these after they land because catching them before would mean distrusting the things we've spent two decades learning to rely on.

115
00:11:09.163 --> 00:11:12.461
[Nina] That's an uncomfortable question for any architecture review.

116
00:11:12.711 --> 00:11:18.933
[Max] The question he leaves teams with: which of the things you trust to tell you you're safe can be turned against you?

117
00:11:22.333 --> 00:11:28.805
[Max] That's Context Window for the week of May fourth. May the fourth be with you.

118
00:11:28.805 --> 00:11:34.845
The worm signed its commits as an AI assistant. Your integrity monitor reported clean.

119
00:11:34.845 --> 00:11:38.727
And the green checkmark was genuine. Stay secure, everyone.

120
00:11:38.977 --> 00:11:45.293
[Nina] If your CI pipeline starts committing poetry, maybe check who's authoring it. See you next week.

121
00:11:45.543 --> 00:11:52.657
[Max] Everything is at contextwindow-sec dot com. The newsletter, full show notes, every source link.

122
00:11:52.657 --> 00:12:03.330
If this episode was useful, share it with someone on your team or leave us a five-star rating wherever you listen.

