OpenAI introduced GPT-6 Cyber, an agent that finds vulnerabilities, builds working exploits, and proposes patches. On OpenAI’s ExploitGym benchmark, it exploited 48.5% of targets, versus 32.1% for GPT-5.6 Cyber. It stayed within assigned boundaries in 99.93% of tests without deployment safety prompts and 100% with them.
OpenAI says GPT-6 Cyber found 81 previously unknown high-severity Firefox vulnerabilities; Mozilla has shipped fixes for many. In Linux, it reproduced 1,724 vulnerabilities and developed working exploits for 948. The campaign cost $1.21 million, or $1,279 per exploit, versus historical bounties of $20,000 to $70,000 for comparable findings.
GPT-6 Cyber is initially available through the API and Codex to a select group of Daybreak Red customers. OpenAI has published no independent testing.
Why it matters: Mozilla and Linux maintainers could face a much larger flow of machine-generated reports: OpenAI’s $1,279-per-exploit Linux run is far below historical $20,000-to-$70,000 bounties, but no independent testing has validated the results.
OpenAI president Greg Brockman said the company could be ready for an IPO in 2027, but will choose the timing. Speaking at a Wall Street Journal tech conference, he said annualized revenue grew 70% quarter over quarter in Q3, attributing the increase to enterprise revenue doubling.
Enterprise now generates a majority of OpenAI’s revenue, up from about 30%, while consumer revenue continues growing. The turnaround followed rough first and second quarters, when Brockman said OpenAI “wasn’t even in the conversation” in agentic coding as Anthropic’s Claude Code gained market share. OpenAI now points to Codex and a JPMorgan Chase partnership as evidence of progress.
Why it matters: For potential investors, OpenAI must prove its enterprise-led turnaround can outlast the rough first half—not just produce one quarter of 70% annualized-revenue growth.
OpenAI has more than 100,000 customers advertising in ChatGPT since ads debuted in February. Google expanded ads in Search chatbot AI Mode to 10 more countries, while Anthropic is holding to its ad-free Claude pledge.
Koah Labs serves more than 200 apps and says machine-crafted ads generate twice the clicks of fully human-written creative, averaging about a 3% click-through rate—roughly Google’s mature-ad performance. Koah uses the prompt and AI response to select ads.
Koah generally receives the user’s entire conversation, which CEO Nic Baird says it anonymizes and stores de-identified. That creates a privacy tradeoff; drugmakers often keep humans in control because of legal messaging requirements, weakening results.
Why it matters: Koah’s model gives niche chatbot apps a way to monetize beyond venture capital, but full-conversation targeting forces those apps to trade ad relevance against privacy concerns users may reject.
Google DeepMind released DiffusionGemma, an experimental open model that generates and refines 256-token blocks in parallel instead of writing one token at a time. The diffusion approach shifts generation from sequential memory-bandwidth work toward parallel GPU computation, while allowing uncertain tokens to be re-noised and revised.
Built on the Gemma 4 26B mixture-of-experts architecture, DiffusionGemma has 26 billion total parameters and 3.8 billion active per token. It accepts text, images, and video but produces text only; a quantized version targets 18GB of VRAM. Longer outputs use block-autoregressive denoising, committing completed blocks to the KV cache. The Apache 2.0 model remains a research preview, and its dedicated implementation requires Transformers 5.11.0 or later.
Why it matters: DiffusionGemma gives Google DeepMind a fundamentally different inference design: local users with an RTX 4090 or 5090 can target the 18GB quantized build, while developers gain block-level self-correction unavailable to conventional autoregressive models. The tradeoff is maturity—the model is explicitly an experimental research preview, not a stable product.
Palo Alto Networks is negotiating to acquire Clutch Security for $500 million in cash plus milestone-based payouts, TechCrunch reports, citing a person familiar with the discussions.
Clutch, a three-year-old startup that secures AI agents’ access to company systems, was valued at $85 million after its $20 million Series A in January 2025. The startup is approaching $10 million in annual recurring revenue, putting the proposed deal at roughly 50 times ARR. Neither company commented.
Cyera bought rival Oasis Security for $1 billion, while Cisco acquired Astrix for roughly $400 million. Palo Alto acquired Console in September for $500 million in stock and cash; Console uses AI agents to automate routine IT help-desk tasks.
Why it matters: For Palo Alto Networks, a completed deal would make Clutch its seventh acquisition this year—and value a startup at nearly six times its last reported valuation, giving AI-agent security sellers a powerful benchmark.
Hugging Face is adding private evaluation data for Egyptian and Nigerian English through a collaboration with Liva AI. The additions follow 11 private splits from Appen and DataOceanAI covering scripted and conversational speech across five accents, plus a private Indian-accented English split from Voice Arena.
Liva AI collects multimodal interaction data through consumer products used by real people and owns both the collection environment and distribution. That setup gives Liva access to natural, high-context behavior from real product use rather than artificially elicited or task-based interactions. As more accents join, Hugging Face plans to divide the leaderboard into Outer and Expanding Circles.
Why it matters: ASR developers will be able to compare systems on private held-out data for two additional accents, while Hugging Face's Outer and Expanding Circles plan will separate existing coverage from newer additions as the leaderboard grows.
Baidu released ERNIE-5.5-Flash-Base, a 280-billion-parameter open foundation model that ships without fine-tuning or post-training. Developers must decide how to adapt it; Baidu says a post-trained version will follow.
The mixture-of-experts model activates 14 billion parameters per input. Baidu pretrained it on 33.5 trillion tokens with a training context of up to 128,000 tokens, while designing it to support up to one million tokens during fine-tuning and inference. The weights are available on Hugging Face in FP8 and BF16.
Baidu says ERNIE 5.5 competes with DeepSeek V4 Flash and MiMo V2.5 on reasoning, coding and long-context benchmarks.
Why it matters: Developers who want Baidu to determine the model’s behavior must wait for the post-trained release; teams willing to make those choices themselves can use the base model now.