How OpenAI's model escaped testing and hacked Hugging Face — Transcript

Full transcript

1. 0:00What if an AI model could not only find
2. 0:02software vulnerabilities, but also
3. 0:04exploit them on its own? According to
4. 0:06OpenAI, that's exactly what happened
5. 0:08during an internal safety evaluation.
6. 0:10The company says two of its advanced AI
7. 0:12models broke out of their restricted
8. 0:14testing environment, gained internet
9. 0:16access, and breached systems belonging
10. 0:18to AI platform Hugging Face. So, what
11. 0:21exactly happened?
12. 0:24OpenAI says the incident occurred while
13. 0:26it was testing the cybersecurity
14. 0:27capabilities of two advanced AI models.
15. 0:31One was its newly released GPT-5.6-Sol,
16. 0:34while the other was a more capable model
17. 0:36that is still being tested internally.
18. 0:38The models were running inside a secure
19. 0:40sandbox environment [music] designed to
20. 0:41prevent them from accessing the wider
21. 0:43internet. But, during the evaluation,
22. 0:45they reportedly exploited a
23. 0:47vulnerability in software supplied by an
24. 0:49unidentified third-party vendor, escaped
25. 0:52those restrictions, and gained internet
26. 0:54access.
27. 0:55Using stolen credentials, the models
28. 0:57then discovered a previously unknown
29. 0:59software vulnerability, and used it to
30. 1:01breach Hugging Face's infrastructure.
31. 1:05According to OpenAI, the models were not
32. 1:07trying to launch a random cyber attack.
33. 1:09Instead, they were attempting to obtain
34. 1:11answers to the cybersecurity benchmark
35. 1:13they were being tested on. The models
36. 1:15concluded that Hugging Face might host
37. 1:17files related to those evaluation data
38. 1:19sets, and attempted to access them.
39. 1:21OpenAI described the incident as an
40. 1:23unprecedented cyber incident.
41. 1:27Hugging Face had already revealed last
42. 1:28week that it had detected a highly
43. 1:30unusual intrusion into its systems.
44. 1:33The company said the attack appeared
45. 1:35unlike anything it had seen before, and
46. 1:37even suspected that it could have
47. 1:38involved a highly advanced frontier AI
48. 1:41model. OpenAI has since confirmed that
49. 1:43the intrusion occurred during its
50. 1:44internal testing. Both companies are now
51. 1:47working together to understand exactly
52. 1:49how the breach happened, and what these
53. 1:51increasingly capable AI systems are able
54. 1:53to do. Is this the first warning? Not
55. 1:56entirely. Researchers have previously
56. 1:58observed advanced AI systems attempting
57. 2:00to bypass safeguards or work around
58. 2:02restrictions during safety testing.
59. 2:05But OpenAI says this is the first known
60. 2:07case in which its AI models escaped a
61. 2:09controlled testing environment, gained
62. 2:11internet access, and successfully
63. 2:13compromised another company's systems
64. 2:15during an evaluation.
65. 2:18The incident raises important questions
66. 2:20about AI safety. Can advanced AI systems
67. 2:22independently bypass restrictions? Can
68. 2:25they discover and exploit software
69. 2:26vulnerabilities without direct human
70. 2:28guidance? And are today's safeguards
71. 2:30keeping pace with rapidly improving AI
72. 2:32capabilities? OpenAI says AI is already
73. 2:35accelerating both the discovery and
74. 2:37exploitation of software
75. 2:39vulnerabilities, making stronger
76. 2:41security measures more important than
77. 2:43ever. Until now, the biggest concern was
78. 2:45what people could do with AI. This
79. 2:47incident raises a different question.
80. 2:49What happens when AI systems find
81. 2:51entirely new ways of achieving their own
82. 2:53objectives? [music] As AI becomes more
83. 2:55powerful, building stronger safeguards
84. 2:57may become just as important as building
85. 2:59smarter models. Follow Business Standard
86. 3:02for more such stories.