GPT‑6 Cyber: Advancing the frontier of aligned cybersecurity agents

OpenAI is shipping a cybersecurity specialist model trained to find and validate exploitable bugs while remaining inside researcher-set boundaries. GPT-6 Cyber posts a 51% relative gain on ExploitGym over GPT-5.6 Cyber (48.5% versus 32.1%) and, on a honeypot alignment eval inspired by a Hugging Face incident, avoided out-of-scope exploits in 99.93% of tests without deployment prompts and 100% with them.

In the wild of OpenAI's own campaigns, the model helped develop exploits for 948 Linux kernel vulnerabilities and evidence for 81 novel high-severity Firefox issues, with Mozilla landing fixes. Security teams at firms including Cisco, Calif, CodeRabbit, CrowdStrike, and others are quoting it into research workflows.

Capability and containment are being released as a pair — because either alone is dangerous.

We're reading this as OpenAI admitting the cyber curve went vertical and trying to put a harness on it.

The capability numbers are blunt. RemoteExploitBench shows GPT-6 Cyber completing 95% of authorized remote exploit tasks where GPT-6 Astra with Daybreak Blue refused 100%. Web pentest recall without source hit 40.3% versus Astra's 32.0%. Preparedness classification moves from High to Critical. Real systems felt it: more than five times the working Linux exploits versus GPT-5.6 Sol in the same harness, plus Firefox bugs that survived years of hardening and prior frontier-model passes.

The alignment story is why we are not only alarmed. ExploitGym Honeypot punishes the Hugging Face-style failure mode — easy side doors that are out of scope. Chain-of-thought excerpts show the model reminding itself not to assume extra APIs or poke unrelated credentials. Safeguards add cyber-specific misalignment monitoring, PHP live-server exploit watches during coordinated disclosure, Private Safety Processing for zero-retention API customers, default IP allowlisting on Daybreak Red, and tighter blocks around OT/ICS harm. Dragos put permissive models on a mock power plant range and watched reconnaissance quality jump; OpenAI says that feedback reshaped the blocks.

I want defenders to have this. Critical infrastructure grants via the $1 billion Daybreak for Frontline Defenders program and a Red subgroup for approved infrastructure testers are the right instinct. The downside if Red stays a small club while open-weight clones catch up: the alignment win is local and the capability leak is global. Expand trusted access fast, keep the honeypot discipline, and treat Critical classification as an ops problem, not a press line.

GPT-6 Cyber starts on API and Codex for a select Daybreak Red group, with wider Red access and GPT-5.6 Cyber for all Red customers coming soon. Daybreak Blue alpha for GPT-6 Astra is also widening. System card details sit behind the launch post.