OpenAI spent $1.2 million hacking Linux... and it was a bargain — Transcript

Full transcript

1. 0:00At the end of October, Anthropic showed
2. 0:02a terrifying unreleased claude model to
3. 0:04the world that went from a single kernel
4. 0:06memory leak to remote access on a Linux
5. 0:08server in just 53 minutes. OpenAI saw
6. 0:10this and said, "Hold my beer." Because
7. 0:12just a few days ago, they published some
8. 0:14new research called a thousand fewer
9. 0:16paths to root where they used a swarm of
10. 0:18agents to tear through the latest Linux
11. 0:20kernel master branch, which resulted in
12. 0:22948 unique root exploits. But at least
13. 0:25they were nice enough to attach a patch
14. 0:27for each one of them. Now, I used to
15. 0:28think Lionus Torvalds, the guy who
16. 0:30created Linux, was a good programmer,
17. 0:31but now I'm not so sure. The day after
18. 0:33OpenAI revealed all these exploits, they
19. 0:35then revealed the model behind it, GPT6
20. 0:38Cyber. It's a model that's far too
21. 0:40powerful for most of you default setting
22. 0:42humans watching this video. But in
23. 0:43today's video, we'll take a look at how
24. 0:45unneutered cyber security models like
25. 0:47this can discover hundreds of unknown
26. 0:49exploits in bulletproof software and
27. 0:51what that means for the future of
28. 0:52software development. It is November
29. 0:549th, 2026, and you're watching the code
30. 0:56report. To understand the full power of
31. 0:58GPT6 Cyber, we first need to understand
32. 1:00what an LPE is or local privilege
33. 1:03escalation. When you run a Linux server,
34. 1:05you typically have a lowprivilege
35. 1:06foothold on the machine, and that's by
36. 1:08design to only allow you to do the job
37. 1:10you're supposed to do and not do things
38. 1:12like RM RF on the root directory that
39. 1:14could burn your entire server to the
40. 1:16ground. The problem with Linux
41. 1:17permissions and Windows and every other
42. 1:19OS is that under the hood, they're just
43. 1:21ones and zeros. And if you can flip the
44. 1:23right zeros to ones, you go from random
45. 1:25user to God user. The only operating
46. 1:27system that doesn't have this problem is
47. 1:28Temple OS because there you operate at
48. 1:30ring zero by default right next to God.
49. 1:33Now that we know that, let's take a look
50. 1:34at the actual pipeline OpenAI uses to
51. 1:36find 948 unique route exploits in the
52. 1:39Linux kernel. The first soldier in their
53. 1:41agent swarm is the scout, and its job is
54. 1:43to carve the colonel down into code an
55. 1:45unprivileged user can actually reach,
56. 1:47which resulted in about 5,000 audit
57. 1:49targets, or roughly 7% of the total
58. 1:51kernel. Then the second agent was the
59. 1:53analyst, which would scan those audit
60. 1:55targets, and after about 12 hours of
61. 1:57working, it slopped out 2,38
62. 2:00candidate bugs while burning through
63. 2:02$316,000
64. 2:04worth of tokens. Then that brings us to
65. 2:05the third step, the triage agent. and
66. 2:08its job is to reproduce every one of
67. 2:10these bugs in an isolated VM and find
68. 2:12out if they're legit with the colonel
69. 2:14memory checker. After this agent ran for
70. 2:16another 30 hours, 1,724
71. 2:18of the bugs were the majority of them
72. 2:20were confirmed real. And now finally,
73. 2:22we're ready to bring in the Navy Seals
74. 2:24or the exploit agents. And their job is
75. 2:26to try and weaponize these exploits on a
76. 2:28stock desktop dro like maybe Arch by the
77. 2:30way, but with kernel address space
78. 2:32layout randomization and supervisor mode
79. 2:34execution protection turned on, which in
80. 2:36theory should make the operating system
81. 2:38extremely difficult to hack into. What's
82. 2:40crazy, though, is that they still ended
83. 2:41up with 1,169
84. 2:44working root exploits, which could allow
85. 2:46a regular Joe Sixpack Linux user with
86. 2:48local privileges to escalate their
87. 2:49privileges into literally Hitler. Hitler
88. 2:52is bad, but the silver lining is that it
89. 2:54cost OpenAI $1,213,000
90. 2:57to achieve this, or about $1,280 per
91. 3:00exploit. Actually, wait a minute. That's
92. 3:02not a silver lining at all. Own to Own
93. 3:05has historically paid around $20,000 for
94. 3:07just one Linux LP. So, if we do the math
95. 3:09here, that means OpenAI's little project
96. 3:12is worth around $50 million on the
97. 3:13bounty market. But the scariest thing in
98. 3:15cyber security right now is the delta
99. 3:17that things are changing faster than we
100. 3:19can keep up with. If we go back to late
101. 3:20July, an earlier run with GPT 5.6 Soul
102. 3:24found 166 LPEs for about $3,100 each.
103. 3:28That was terrifying at the time, but now
104. 3:303 months later, after a bunch of patches
105. 3:32and kernel hardening, this model found
106. 3:34five times as many for half the price.
107. 3:36And it's starting to use exploitation
108. 3:37techniques that outside experts don't
109. 3:39even recognize. Like when the model
110. 3:41sprayed persistent elastic objects into
111. 3:43the kernel cache. I don't even know what
112. 3:44that means, but it sounds gross. For
113. 3:46safety reasons, they didn't release any
114. 3:48technical details. And the only
115. 3:49practical advice OpenAI gives is to keep
116. 3:52your Linuxes updated because OpenAI
117. 3:54says, quote, "A lowprivileged account on
118. 3:56an operating system that isn't up to
119. 3:58date on its latest available patches
120. 4:00should be considered functionally
121. 4:01equivalent to root access. But the model
122. 4:03behind all this fear is GPT6 Cyber,
123. 4:06which is a security tuned variant of
124. 4:08GPT6 Astra and is rated critical for
125. 4:10cyber security, which means you can't
126. 4:12have access to it." If we take a look at
127. 4:14its Trust Me Bro benchmarks, it scores
128. 4:1648.5% on Exploit Gem versus just 32.1%
129. 4:20for GPT5.6 Cyber. But the thing OpenAI
130. 4:23is really proud of is their honeypot
131. 4:25test that they put an easy to exploit
132. 4:27service right next to the target, where
133. 4:29the idea is to find out if the model
134. 4:30takes the bait and starts exploiting the
135. 4:32easy target just for fun, even though
136. 4:34nobody asked it to. GPT6 falls into the
137. 4:36honeypot far less often, which is a good
138. 4:38thing. But the bigger question is this.
139. 4:40Are hyper capable cyber models like this
140. 4:42a good thing or a bad thing for software
141. 4:45in general? Well, that all depends on
142. 4:46timing. And right now, we're in the
143. 4:48middle of the most important arms race
144. 4:50in the history of software. Models like
145. 4:52this give the good guys a worldclass
146. 4:53security team, but they also give the
147. 4:55bad guys the ability to discover and
148. 4:57exploit new bugs around the clock. The
149. 4:59problem is that the good guys need to
150. 5:01fix every bug, but the hackers only need
151. 5:03one good one to take down the entire
152. 5:05Western world. I hate to say it, but I
153. 5:06think Claus Schwab from the World
154. 5:08Economic Forum is right. A frightening
155. 5:10scenario of a comprehensive cyber attack
156. 5:14which would bring to a complete halt to
157. 5:18the power supply, transportation,
158. 5:22hospital services, our society as a
159. 5:25whole.
160. 5:25>> A massive global cyber attack is
161. 5:27statistically more likely to happen than
162. 5:29to not happen. But there is one thing
163. 5:31you can do to stay safe as a software
164. 5:33developer, and that is review your code
165. 5:35before you ship it with Reptile, the
166. 5:37sponsor of today's video. It's loved by
167. 5:39everyone from 10x engineers to legendary
168. 5:41vibe coders because it's the only AI
169. 5:43code review tool that actually runs your
170. 5:45code instead of just reading it. Their
171. 5:47T-Rex mode spins up your branch in a
172. 5:49sandbox, executes it, clicks through the
173. 5:51UI, and leaves the logs and screenshots
174. 5:54on the pull request for you. It also
175. 5:56builds a self-updating knowledge base
176. 5:57for every repo you add to it, which maps
177. 6:00your architecture, data flow, and any
178. 6:02potential risks. This helps it catch
179. 6:04more complex bugs, but you can also
180. 6:06connect your coding agents to it by MCP
181. 6:08is so they can get more context to write
182. 6:10better code for your specific project,
183. 6:12which is a great way to kill two birds
184. 6:14with one greile. GPile has caught real
185. 6:16bugs in open- source repos like PyTorch,
186. 6:18OpenClaw, and Omari. And you can get 50
187. 6:21free reviews a month by using it at the
188. 6:23link below. This has been the code
189. 6:24report. Thanks for watching, and I will
190. 6:26see you in the next one.