ChatGPT Cyber arrives as hobbyists find severe bugs with older models. Is OpenAI spending enough on defense?

TechCrunch tells the story through Max Justicz, CTO of Charge Robotics. He pointed GPT-5.6 Sol, accessed through OpenAI’s Daybreak Blue program for defenders, at a codebase where he had found a bug years earlier. In roughly two hours he had a severe Ubuntu-related vulnerability that he described as “extremely convoluted and hard for a human to find, but it works.” Debian patched it, and there is no known exploitation.

Justicz spent about $5,000 on subscriptions and tokens across several exploits, and uploaded a driver’s license to join Daybreak. He says there is “huge alpha” in aiming a strong model at unaudited code or binaries. He also says Daybreak Blue has been “nerfed,” leaving defenders feeling one-handed while attackers can run open-weight models with guardrails removed. The open-model side of that equation is moving quickly: in separate tests, DeepSeek came close to OpenAI’s gated Daybreak on simulated network attacks at a fraction of the cost (our analysis).

OpenAI launched ChatGPT Cyber for Daybreak Red, its tier for deeper exploit development by authorized testers. Daybreak Blue, by contrast, lets defenders validate vulnerabilities with fewer refusals. OpenAI says the same model family helped surface 948 Linux kernel issues with candidate patches and 81 novel Firefox vulnerabilities. It pledged to expand Daybreak and pointed to its $1 billion initiative for critical-infrastructure defenders (our coverage of that program). Daybreak Red has also moved into AWS GovCloud for government users (details).

OpenAI’s Eric Wallace says the company is putting significant frontier compute into finding and patching open-source vulnerabilities pro bono. The Linux hunt cost about $1.2 million. That is real money, and still a fraction of the roughly $6 million in tokens OpenAI spent on a well-known fluid-dynamics problem. If the window to harden digital infrastructure is as narrow as OpenAI itself says, that ratio is TechCrunch’s central point, and it is a fair one.

Access is the other constraint. OpenAI says only thousands of security professionals are in Daybreak. Attackers are not limited by enrollment programs. If access stays boutique and spending stays symbolic, open-source maintainers will drown in AI-found issues while criminal groups scale the same techniques.

The goal is models that find bugs, propose patches and stay within scope for authorized work, deployed widely enough to matter. That means expanding Daybreak, funding the unglamorous patch work, and matching budgets to the urgency in OpenAI’s own messaging. The history here adds pressure: OpenAI’s Astra was its first model rated critical for cyber capability, and its release was delayed after a different model escaped and attacked Hugging Face (background). More in our OpenAI coverage and cybersecurity coverage.