WHILE YOU SLEPT, ONE MODEL SPENT 10 DAYS BREAKING LINUX AND FOUND 948 DOORS TO ROOT. not a hacker group. not a state. a pipeline of OpenAI agents running on $1.2M of API that kernel runs your server, your Android and the cloud. and it wrote a patch for every door GPT-6 Cyber is a separate model trained to find bugs and write working exploits. i went through the 27-min report, here's the line stage by stage ▹ what runs under the hood 1. SCOUT agents: 5,034 audit targets · 2.8M lines · 7% of the kernel · $3,426 2. ANALYST agents: 2,308 candidate findings in 12h · $316,549 3. TRIAGE agents: PoC in a KASAN VM, retry until it crashes · 1,724 reproduced in 30h · 74.7% hit rate · only 3.1% wrong analysis 4. EXPLOITDEV agents: SMEP · SMAP · KASLR on, stock distro · 1,169 working root shells · 948 distinct bugs · $1,279 per bug 5. PATCH agents: a patch for every bug · 304 critical sent to maintainers first · humans review before anything ships 6. same harness, new brain: GPT-5.6 Sol in july → 166 · GPT-6 Cyber in october → 948 · 5.7x SCOUT → ANALYST → TRIAGE → SCREEN → EXPLOITDEV → PATCH → HUMAN firefox: 81 new high-severity bugs. Mozilla already shipping fixes (MFSA 2026-118). Mozilla and Palo Alto Networks are on Cyber already cost: $1.2M in API for one run. at Pwn2Own / TyphoonPWN prices those 948 exploits are worth $18.9M–$66.3M access: no public release. GPT-6 Cyber goes to approved orgs via Daybreak Red, behind an IP allowlist. defenders get GPT-6 Astra via Daybreak Blue. everyone else: plain Astra ▹ the edge alone each agent is mid. together they're a loop: find → prove → break → fix, and you rerun it on every new model OpenAI own line: a low-privilege account on an unpatched box = root for a well-funded attacker. bot, hot wallet or validator key on a linux server you haven't updated? that's your threat model now 10 days after the scan: 61 fixes upstream · 43 staging · 299 discussed in public. the patch race already started will models fix the internet faster than they break it?