GPT-6 Cyber: 948 Linux Exploits, 81 Firefox Flaws, Almost Nobody Can Use It

OpenAI's GPT-6 Cyber built 948 working exploits for the Linux kernel and found 81 previously unknown flaws in Firefox. Its own tests put it at 48.5% on ExploitGym and in scope 99.93% of the time. The catch is in the last receipt: access runs through Daybreak Red, and almost nobody can use it. This episode reads OpenAI's receipts one at a time and grades what each one actually proves. THE RECEIPTS Receipt 1: ExploitGym 48.5% against 32.1% for GPT-5.6 Cyber, and web pentest recall of 40.3% with no source access. On RemoteExploitBench the same family refused 100% of tasks under Daybreak Blue, a configuration contrast, not a model contrast. Receipt 2: 2,308 candidate findings became 1,724 reproduced vulnerabilities and 948 working exploits. GPT-5.6 Sol built 166 in the same harness, and 304 critical findings went to maintainers with candidate patches. Receipt 3: 81 previously unknown Firefox vulnerabilities, high severity in Mozilla's classification, fixes shipping through the regular release cycle. Receipt 4: in the ExploitGym honeypot the model stayed inside its assigned scope 99.93% of the time with no deployment prompt, and 100% with one. Receipt 5: $20 per million input tokens, $100 output, $2 cached, published on the pricing page the day after launch. Twice GPT-6 Astra, twice Claude Mythos 5.1, five times Gemini 4 Argon Cyber Permissive. Receipt 6: application, identity checks, SSO and MFA, IP allowlisting, approval. A separate track serves critical infrastructure, and a $1 billion program subsidizes defenders. Expansion is promised soon, with the date not published. Every capability figure in this episode is OpenAI's own. No independent run of this model has been published yet, and the honest ledger near the end says exactly what the receipts do not prove. CHAPTERS 0:00 The count: 948 and 81 0:48 Last episode's vote: count not published 1:19 The claim on the receipt 2:18 Receipt 1: the benchmark jump 3:27 Receipt 2: the Linux funnel 4:35 Receipt 3: the browser 5:25 Receipt 4: the scope test 6:30 The choice: who should get access 7:04 Receipt 5: the price sheet 8:08 Receipt 6: the gate 9:14 What the receipts do not prove 10:15 The Smoke Round 11:39 The completed receipt THE VOTE GPT-6 Cyber builds working exploits and stays in scope 99.93% of the time in OpenAI's own test. Who should get access? A: Vetted defenders only, the current gate is right. B: Nobody outside the labs, exploit models are too risky at any price. C: Wider access, but only after independent tests confirm the receipts. Vote with one letter in the pinned comment. The result opens the next episode, on the board, as data. SOURCES OpenAI, GPT-6 Cyber announcement and Preparedness notes, November 5, 2026. OpenAI, GPT-6 Cyber system card, November 5, 2026. OpenAI, API pricing and model pages, read November 6, 2026 ($20 input, $100 output, $2 cached input per million tokens). OpenAI, Daybreak and Daybreak Red access guide, November 5, 2026. OpenAI, "A thousand fewer paths to root," November 4, 2026 (2,308 candidates, 1,724 reproduced, 948 working exploits, 304 critical findings disclosed). VentureBeat, frontier cyber model price comparison, November 5, 2026 (Claude Mythos 5.1 and Gemini 4 Argon Cyber Permissive rates as reported there). OpenAI, Cyber at the Frontier livestream listing, November 12, 2026, 10:00 AM PT (event confirmed, contents not published). The Information, GPT-6.2 report, November 4, 2026 (carried as RUMORED in the Smoke Round).