OpenAI in early talks to buy security firm Trail of Bits
OpenAI's security ambitions got a name attached to them today, Google opened its frontier models to European scientists, and a Chinese lab put an agent inside a phone and sold it by the token plan.
OpenAI is in early talks to acquire Trail of Bits, the New York security firm that has spent fourteen years finding and fixing vulnerabilities in open-source software. The Information reported the discussions Tuesday, citing people familiar with the matter; the deal size is undisclosed, the talks are described as early and could fall apart, and neither company has commented. No second outlet has independently confirmed the talks — every English-language write-up so far traces back to that one report, so treat the price and the timing as open questions rather than facts.
The fit is not hard to see. Trail of Bits has been inside OpenAI's orbit since June, when the two ran "Patch the Planet," a program that pointed GPT-5.5-Cyber and Codex Security at open-source projects in Python, cURL and Go and had the firm's engineers review and merge the patches. Chief executive Dan Guido's line at the time — that Codex let his team deliver in a day what used to take weeks — is the whole business case in one sentence. The firm also has the credentials to back it up: it finished second in DARPA's AI Cyber Challenge, winning $3 million with a system that found 28 vulnerabilities, applied 19 patches across 20 weakness classes, and did it at a reported $181 of compute per scoring point. That last number is the interesting one. It means world-class automated vulnerability discovery does not require a frontier-scale budget, which is exactly the kind of claim an AI lab selling cyber capability to enterprises wants to own.
Why it matters: security is becoming a product line for the labs, not a research paper. OpenAI and Anthropic are both positioning to sell models that find and fix bugs to large customers, while deliberately dulling the same capability in the models the public gets. Buying the most trusted name in open-source auditing would hand OpenAI the scarce asset in that market — engineers with a decade of standing in the open-source community — on top of the models. We have watched the adjacent move already, when Cloudflare turned OpenAI's cyber model into an edge patch shop; a lab owning the firm that does the auditing is the next step in the same direction.
Google DeepMind and Google Cloud will give up to 200 European scientific researchers access to frontier models and $3 million in compute credits. The pledge, announced at an AI summit in Dublin, runs through the European Union's RAISE program, with the researchers chosen by the European Commission rather than by Google — a detail the Reuters wire downplayed and the company's own blog post led with. The toolset on offer is the applied end of DeepMind's catalog: AlphaFold 3 for protein structure, AlphaGenome, WeatherNext for forecasting and AlphaEvolve.
The take: $3 million is a rounding error against European compute spending, which is precisely why the framing matters more than the money. Google is paying in access and goodwill while Brussels writes the rules that govern it, and the recipients are selected by the Commission — a structure that makes the grant look less like charity and more like a stake in the continent's science policy. Europe's problem has never been researchers; it has been compute and models it does not own.
StepFun's first agent phone is on sale, and it ships with a token plan attached. The STEPX Neo opened pre-orders at ¥3,999 (roughly $560) in China, bundled with a twelve-month plan worth 10 billion tokens a month, with general sales from October 16. Internally it runs Step AOS, which StepFun calls an agent-native operating system: the assistant can make system-level tool calls, chain skills and MCP tools the way a desktop coding agent does, and shows its work on a small screen on the back of the phone. Permissions are gated through a dedicated agent-security panel that logs every tool call and requires a confirmation before anything is deleted, paid for or moved between apps.
The take: substance is in the ecosystem, not the hardware specs. StepFun says partners like Alipay, WeChat, Didi and Meituan expose their services to the agent directly, which skips the visual-recognition-and-mock-taps workaround that makes phone agents brittle. There is no announced international availability and no global pricing, so read this as a China-market data point — but it is the same packaging bet ByteDance made when its agent phone went on sale in September and gave apps a veto: sell the agent and the tokens as one product, and the phone becomes a subscription with a screen.
What to watch: whether OpenAI confirms or denies the Trail of Bits talks inside a week — an unconfirmed acquisition rumor that survives that long usually becomes a real one — and whether the agent-phone token bundle shows up in a Western launch.
Would you buy a phone that comes with a monthly token allowance, or does that just make it a subscription you can drop? Tell us in the comments.
