Houthis Used Claude To Develop Weapons. What About AI Safeguards? | The Palki Sharma Show | IGR — Transcript

Full transcript

1. 0:02In 2014, the Houthis were a rag tag
2. 0:05militia. They had limited military
3. 0:07equipment. They were fighting battles in
4. 0:09sandals and they were seen as an
5. 0:11unsophisticated fighting force. 12 years
6. 0:15later, they are a formidable force.
7. 0:18Today, the Houthis launch missiles, they
8. 0:20use drones, and it turns out they're
9. 0:22using AI to make those drones,
10. 0:25especially Clawude. The Houthis are said
11. 0:28to be using clawude to make drones.
12. 0:31Clearly, they did not settle for the
13. 0:32free trial. But how did they manage
14. 0:35this? What happened to AI safeguards?
15. 0:38And what does this mean for warfare?
16. 0:41Let's start with Anthropic. That's the
17. 0:42company behind AI model Claude. It has
18. 0:45published a new report, a threat
19. 0:48intelligence report. It's called
20. 0:51detecting and countering misuse of AI
21. 0:54September 2026. That's the title of the
22. 0:56report. As the name suggests, the report
23. 0:58talks about the misuse of AI. And
24. 1:01there's one example that Anthropic has
25. 1:04highlighted. A cell of threat actors in
26. 1:07northern Yemen. That's what the report
27. 1:09has said. A cell of threat actors in
28. 1:11northern Yemen. They've not named the
29. 1:13Houthis, but they did not have to. The
30. 1:15inference is obvious. The Houthis
31. 1:17control northern Yemen, and they are
32. 1:18waging a war. So, the report clearly
33. 1:21talks about the Houthis.
34. 1:24And here's what it says. The Houthis
35. 1:27used claude for their for three weapons
36. 1:29projects. One involved a guided rocket.
37. 1:32The second involved a ballistic missile.
38. 1:34And the third was a missile project. It
39. 1:37included a hypersonic glide vehicle.
40. 1:41Now all missiles need software. Software
41. 1:44that helps them navigate and stay on
42. 1:45course. And apparently the Houthis used
43. 1:47Claude to develop this software.
44. 1:51Claude did the coding. It was used to
45. 1:53run simulations to check how it works.
46. 1:56The report says the Houthis used
47. 1:57multiple AI sessions at once. One worked
48. 2:00on code. Another did the research and a
49. 2:04third one reviewed the code.
50. 2:07Looks like the Houthies have aced AI
51. 2:09multitasking. Basically, they used AI as
52. 2:11an engineering assistant.
53. 2:14And then came the real world test. The
54. 2:16Houthis test fired a guided rocket.
55. 2:20According to Anthropic, that test
56. 2:22failed. So the Houthis went back to
57. 2:24Claude to figure out what went wrong.
58. 2:27Clearly they did a lot of brainstorming
59. 2:29with Claude. The question is how was
60. 2:32this not flagged?
61. 2:34Anthropic says it was. It says they
62. 2:37flagged the Houthi requests many times.
63. 2:40They also blocked some accounts, but
64. 2:42some requests still got through.
65. 2:45And that's a big problem because AI is
66. 2:48changing the way wars are fought.
67. 2:51This is not the only instance where it's
68. 2:53been used to infl used in conflict. Look
69. 2:56at Israel's war in Gaza. Reports say
70. 2:58they used AI to identify targets. In the
71. 3:02United States, they have something
72. 3:03called Project Maven. It uses AI to
73. 3:06analyze drone footage, satellite images,
74. 3:08and other intelligence. And in Ukraine
75. 3:11too, AI enabled tools are helping
76. 3:13process drone footage. So AI is playing
77. 3:15an outsized role in conflicts.
78. 3:20What happens when the same technology
79. 3:22falls into the hands of non-state
80. 3:23actors?
81. 3:25Where are the guard rails? Who is
82. 3:26keeping watch?
83. 3:29Anthropic says it flagged the Houthi
84. 3:31link activity, but clearly it was not
85. 3:33enough. The safeguards are not
86. 3:35foolproof. Users can always rephrase
87. 3:38requests. They can break a task into
88. 3:40smaller steps. They can ask
89. 3:41non-conspicuous questions that do not
90. 3:44look dangerous in isolation.
91. 3:47Also, Anthropic seems to have quite the
92. 3:49diverse portfolio of customers. On one
93. 3:52side, they have the Houthis. On the
94. 3:54other, they have the Pentagon. That's
95. 3:56the US Defense Ministry. And the irony
96. 3:58does not end there. Anthropic had a
97. 4:02public dispute, a public falling out
98. 4:04with the US Defense Department, a
99. 4:06dispute over AI usage in wars. The US
100. 4:09military wanted access to advanced AI,
101. 4:13but Anthropic refused. It put limits on
102. 4:16how AI is used in conflict. So, the
103. 4:18company wants guard rails. But the
104. 4:20question is, who's going to enforce it?
105. 4:24Is it the companies that build AI? Is it
106. 4:26the governments that buy AI? Or is it
107. 4:29the countries that try to regulate it?
108. 4:32Because everyone wants the advantage
109. 4:34that AI can offer, but nobody seems to
110. 4:36want the responsibility that comes with
111. 4:38it.
112. 4:39If a militia can use a commercial AI
113. 4:41tool to develop weapons, this is no
114. 4:44longer just a tech question.
115. 4:46It's a question of who gets access to
116. 4:48potentially dangerous technology and who
117. 4:51pays the price for it.