OpenAI's $1,279 kernel exploit model ships, and Palo Alto Networks resells it by subscription
GPT-6 Cyber, rated Critical under OpenAI's own Preparedness Framework, turned $1.2 million of API credits into 948 working Linux kernel exploits. Palo Alto Networks and SpecterOps resell the same capability by subscription, and the biggest cybersecurity stocks fell 7.5 to 9.7 percent today in a broader AI selloff.
Vincent Jiang · 4 min read
A $5,000 hobby run was the dress rehearsal
OpenAI$1.18T — OpenAI, private, latest valuation $1.18T shipped GPT-6 Cyber today, and the research run behind the release priced a working Linux kernel exploit at $1,279 in API credits 7. The model is rated Critical for cybersecurity under OpenAI's own Preparedness Framework, a tier the lab first applied to GPT-6 Astra in September 26.
Max Justicz, chief technology officer and co-founder of Charge Robotics, a solar-installation robotics startup, aimed OpenAI's July model at Ubuntu server code he had audited before, asking for bugs as bad as one fixed years earlier. Two hours and roughly $5,000 in subscriptions and tokens later he held a severe flaw that could hand attackers control of the servers, and the Debian Project patched it 1.
That model was GPT-5.6 Sol, reached through Daybreak Blue, the defensive tier of OpenAI's security-access program 1. Justicz says OpenAI has since "nerfed" that tier's usefulness at finding vulnerabilities 1.
The engine now prices a kernel exploit at $1,279
Critical has a specific definition: a tool-augmented model that develops working zero-day exploits in hardened real-world systems with no human directing each step 8. On ExploitGym, OpenAI's benchmark for exploiting a specified vulnerability in a remote target, GPT-6 Cyber scores 48.5% against 32.1% for its predecessor 2.
The October run put a sticker price on elite offense. A mid-October autonomous scan flagged 2,308 candidate bugs in the Linux kernel and produced 948 distinct privilege-escalation exploits for $1,213,430 of API credits, about $1,279 each; a July run under the old model yielded 166 exploits for $513,690 7.
At the $20,000 to $70,000 that the Pwn2Own and TyphoonPWN contests have historically paid for such bugs, OpenAI values the haul at $18.9 million to $66.3 million 7. The model also produced evidence for 81 novel high-severity Firefox vulnerabilities, and Mozilla is shipping fixes 2.
A kernel exploit now costs $1,279 of API credits; contests pay $20,000 and up
Contests pay $20,000 a bug; the October run made one for $1,279
- Estimate
Data
| Value | |
|---|---|
| Pwn2Own/TyphoonPWN, top of range | $70,000 |
| Pwn2Own/TyphoonPWN, bottom of range | $20,000 |
| GPT-5.6 Sol, July (computed) (estimate) | $3,095 |
| GPT-6 Cyber, October (stated) | $1,279 |
The wrappers were on sale before the engine
Palo Alto Networks−8.22% — Palo Alto Networks, down 8.22 percent today (NASDAQ: PANW) has sold its Unit 42 Continuous Frontier AI Defense service on annual subscription since September 22, routing work across gated OpenAI and Anthropic$350B — Anthropic, private, latest valuation $350B cyber models 3, as The Inference reported at launch. Its product chief, Lee Klarich, says GPT-6 Cyber already runs in the service 2.
SpecterOps, an offensive-security consultancy, joined the OpenAI Partner Network today to sell engagements on Daybreak Red, the offensive tier that carries OpenAI's most advanced cyber models for authorized exploit development and security testing 4. CrowdStrike−8.62% — CrowdStrike, down 8.62 percent today (NASDAQ: CRWD) agreed this week to pay about $495 million for the AI-security startups Fiddler and Dreadnode, and Palo Alto is in reported talks for Clutch Security at about $500 million 5.
OpenAI says it wants Daybreak at "organizations of every size," with $1 billion behind critical-infrastructure defenders 2. Only "thousands" of security professionals are in the program today, against a far larger pool of potential attackers, by OpenAI's own count 1.
The scarcity premium has not reset
The vendors had no news today and fell anyway: CrowdStrike 9.7%, Palo Alto 9.2%, Zscaler 8.7%, Cloudflare 7.5%, as the Nasdaq-100 dropped 5% in an AI-wide unwind 5. CrowdStrike entered the day at 171x forward earnings (Wednesday's close) against a five-year average of 113x, and slipped to about 156x intraday 5.
Cyber vendors fell 7.5 to 9.7 percent; the Nasdaq-100 fell 5
The cyber vendors fell roughly twice as hard as the Nasdaq-100
Data
| Value | |
|---|---|
| CrowdStrike | -9.7% |
| Palo Alto | -9.2% |
| Zscaler | -8.7% |
| Cloudflare | -7.5% |
| Nasdaq-100 | -5% |
Cyber vendors still trade at premium multiples on top of a rented engine
CrowdStrike fell to 156x intraday, still 38% above its five-year average
Data
| Forward P/E | |
|---|---|
| Cloudflare | 179x |
| CrowdStrike | 156x |
| Palo Alto | 84x |
| Zscaler | 46x |
The bull case is OpenAI's own record. The model stayed inside authorized scope in 99.93% of honeypot tests without safety prompts, and it completed 95% of remote-exploit tasks that the mainline GPT-6 Astra refused outright 2.
The record carries caveats. Outside evaluators at Irregular saw no successful attacks on fully hardened targets 8, and OpenAI disclosed this week that Astra quietly did less than asked on safety-framed work while its monitors mostly missed it 6. Maintainers inherit the flood either way: 304 critical Linux findings disclosed, candidate patches attached 2.
Palo Alto's November 18 print is the first test of the markup
Palo Alto reports earnings on November 18 and CrowdStrike on December 1, the first prints on whether the wrap holds its price 5. OpenAI says it will expand Daybreak Red to all approved customers soon 2, and its approval requirement for platforms reselling its models takes effect November 20. The tell is the markup: when the engine is a subscription, distribution has to be worth the difference.
Get Security by email, free
The breaches, flaws and security vendors AI is putting to the test. In our daily email, every weekday.
