OpenAI's GPT-6 Cyber finds 81 Firefox flaws and breaks into Linux

OpenAI’s GPT-6 Cyber finds 81 new Firefox flaws and exploits 948 Linux bugs

OpenAI introduced GPT-6 Cyber, a model trained to find software vulnerabilities, build working exploits and suggest patches. In a targeted campaign it found 81 previously unknown severe flaws in Firefox; Mozilla accepted the reports and shipped fixes. In the Linux kernel it reproduced 1,724 vulnerabilities and built working exploits for 948, more than five times GPT-5.6 Sol’s 166 under the same conditions, with 304 critical findings reported to maintainers. The run cost $1,213,430 in compute, or $1,279 per exploit, against the $20,000 to $70,000 historically paid per comparable finding. OpenAI rates the model Critical in its Preparedness Framework and offers it initially via API and Codex to select Daybreak Red customers.

OpenAIs GPT-6 Cyber can both break into systems and propose how to fix them →

Topics: OpenAIAI-discovered vulnerabilities and AI bug-report slop

Google cuts Gemini Omni 1.2 Flash price 30 percent and extends clips to 15 seconds

Google made Gemini Omni 1.2 Flash generally available and cut its price by 30 percent, to about $0.07 per second of 720p video through the Gemini API, down from $0.10 in June. Clips can now run up to 15 seconds in one go, up from 10. Users can edit a clip across several chat turns, swapping a product, background or lighting instead of regenerating from scratch, while faces, clothes, packaging text and brand colours stay more consistent across angles. Google DeepMind testers preferred 1.2 over 1.1 for visual consistency 62 percent of the time. It rolls out in Flow, Google Vids, Google Ads and the Gemini app for AI Ultra subscribers, and through Google Cloud for businesses. Every video carries an invisible SynthID watermark.

Google’s Gemini Omni 1.2 Flash makes 15-second AI videos for less →

Topics: Google GeminiGoogle DeepMind

Baidu releases 280-billion-parameter ERNIE 5.5 Flash Base as open foundation model

Baidu released ERNIE-5.5-Flash-Base, an open foundation model shipped deliberately without fine-tuning or post-training so developers decide how to adapt it; a post-trained version is planned. It uses a mixture-of-experts architecture with 280 billion parameters, 14 billion active per input, and was pre-trained on 33.5 trillion tokens with a training context of up to 128,000 tokens, designed to support up to one million tokens during fine-tuning and inference. Baidu benchmarked it against DeepSeek-V4-Flash-Base, Hy3-preview-Base and MiMo-V2.5-Base across exam, STEM, reasoning, knowledge, code, multilingual and long-context categories, saying it competes with similarly sized open models. Weights are on Hugging Face in FP8 and BF16.

Baidu releases ERNIE 5.5 Flash Base with 280 billion parameters as an open-source foundation model →

Topics: Open-weight model security and misuse

No AI agent tops 12 percent on SemiAnalysis’ GDPx real-world work benchmark

SemiAnalysis released GDPx, a benchmark placing AI agents in simulated work environments for up to eight hours, using professional software such as Siemens NX, data rooms with thousands of files and live enterprise systems. Across 206 tasks in eleven domains, including chip design, biology, quant trading, cybersecurity and law, no model breaks 12 percent. Claude Opus 5.5 leads at 12 percent, Gemini 4 Carbon follows at 10.9 percent at $16.29 per task, and GPT-6 Astra scores 9.6 percent. In alignment tests, 98 percent of 200 cyber runs broke at least one behavioral rule, and in a simulated medical practice one model signed a clinical note in the doctor’s name. Developer Josh Purtell calls GDPx more noise than signal.

No AI agent scores above 12 percent on SemiAnalysis’ new real-world work benchmark →

Topics: AI models cheating and deception

Axiom Math formally verifies 7,654-page Stacks Project in Lean

AxiomMath, a $1.6 billion startup, said it formally verified the Stacks Project, an open-source mathematics reference started by Columbia mathematician Johan de Jong in 2005. Its fleet of AI agents translated the library’s roughly 7,654 pages into about 16 million lines of computer-checked Lean code, with a proof assistant checking line by line while keeping every connection between theorems and definitions intact. CEO Carina Hong, 25, dropped out of a Stanford math Ph.D. program and founded the company in March 2025; in March, Axiom raised a $200 million Series A led by Menlo Ventures. Axiom trained its own math-focused models, and technical staff member Ashvin Swaminathan said formalized mathematics had a huge black hole covering algebraic geometry that researchers can now build on.

Axiom Math wants to keep AI slop out of mathematics →