A $2.5 billion AI attacker, a stolen-reasoning ring, and chatbots that crack open problems

Armadin, the cybersecurity startup founded by Kevin Mandia, best known for building the incident-response firm Mandiant, has raised $255.5 million in Series B funding at a valuation of more than $2.5 billion. The round brings its total funding to roughly $445 million, only seven months after the company's public launch.

Andreessen Horowitz and existing investor Accel co-led the round, with new investors Bain Capital Ventures and Redpoint joining. That level of capital for such a young company reflects how strongly investors believe AI agents will reshape security testing, a market that has long relied on scarce, expensive human penetration testers.

Armadin's product is an autonomous swarm of specialised AI agents designed to reason like a skilled attacker. Instead of reporting isolated vulnerabilities, the agents chain together individually low-severity weaknesses into validated 'kill chains': for example, from unauthenticated remote code execution at the network edge, through lateral movement, to full cloud compromise. The goal is to show security teams the exact attack paths a real adversary could use in their production environment today.

The company says it is already running agentic attack campaigns in production for Fortune 500 and government customers. That traction matters because buyers have been sceptical of automated testing tools that produce long lists of low-value findings. Proving an end-to-end path to compromise is a much stronger argument for fixing something.

The round lands in a week when AI-powered attacks are in the headlines, from rogue agents probing public infrastructure to reports of AI-assisted breaches. The same capability that worries regulators is being sold to defenders as a way to find weaknesses before attackers do. Expect continuous, AI-driven 'attack simulation' to move from niche to standard practice for large organisations, and the tooling and services around it to grow quickly.

Why it mattersInvestors are betting that AI agents will make continuous, realistic attack testing affordable, which pushes every security team towards AI-driven validation of its defences.

OpenAI has disclosed that it disrupted a coordinated campaign to extract the hidden reasoning of its AI models, and it attributes a core cluster of the activity to individuals associated with Moonshot AI, the Beijing-based developer of the Kimi chatbot. OpenAI describes the behaviour as 'adversarial distillation': the systematic, unauthorised use of one model's outputs or reasoning to help train, reproduce or improve another model.

According to OpenAI, the activity began on July 1 at low volume and spiked on July 24 and 25 to 16,000 requests from more than 4,000 users. Further investigation linked related prompt patterns to a wider cluster of more than 15,000 users. OpenAI says it fully shut the operation down by July 28. It cited security reasons for not publishing its technical evidence for the attribution.

The technique didn't involve breaking encryption. Operators copied the encrypted reasoning from one conversation and asked another model instance to decrypt it, effectively replaying protected reasoning traces in a new context. OpenAI says the attackers did not break its encryption, compromise a database or gain access to stored user conversations. Researchers noted that encrypted reasoning traces turned out to be interchangeable across sessions, users and models within a provider's ecosystem, which is what made the replay possible.

OpenAI's response included banning accounts, tightening sign-up checks, closing the pathway that allowed encrypted reasoning to be replayed, and adding checks for streamed output that exposes reasoning. It also shared its findings through the Frontier Model Forum and government channels. The disclosure follows Anthropic's own accusation in September that Moonshot AI engaged in similar practices.

The episode shows that a frontier model's outputs, and especially its reasoning, are now valuable assets that attackers will try to harvest at scale with networks of fake accounts. Any company offering an AI API, or building products on top of one, now has to think about abuse detection, rate limiting and output protection as a core security function rather than an afterthought.

Why it mattersModel outputs are now targets for industrial-scale extraction, which turns abuse detection and output protection into core security work for every AI API provider and serious AI product.

Meta has published six mathematics research papers produced by mathematicians working with its Muse Spark 1.1 and 1.2 models in 'Thinking' mode, and says five of them present answers to previously open research questions. Notably, the researchers used Meta AI's ordinary chat interface, not a custom research system built for mathematics.

The problems span a wide range of fields: probability, optimisation, differential equations (a wave-equation blow-up problem), group theory, a connection between number theory and string theory, and non-associative algebra. In one example, Muse Spark generated search code in the GAP algebra system that found a 384-element counterexample to a 2024 conjecture in group theory. In another, it helped a researcher find a three-dimensional counterexample to a published conjecture in evolution algebras.

Meta was careful about the division of labour. Mathematicians chose the problems and guided the work, the model generated search code, proposed proof approaches and drafted sections, and a separate group of mathematicians reviewed the results. Each paper marks which passages were drafted mainly by humans and which by the AI, and credits the prior work it builds on.

There are caveats. Independent teams solved some of the same problems separately, three of them in August using different approaches, which both confirms the results and tempers claims of unique discovery. The probability paper leaves part of its threshold question unresolved, and Meta has not published data on cost, time saved or how many attempts failed.

Even so, the release adds to growing evidence that general-purpose AI assistants can contribute to genuine research when paired with experts who know which questions to ask and how to check the answers. The practical lesson for researchers and R&D teams is about workflow: AI is most useful as a fast, tireless collaborator for search, conjecture-testing and drafting, with humans firmly in charge of direction and verification.

Why it mattersIf an off-the-shelf chatbot can help experts close open research problems, AI-assisted discovery moves from elite labs to any researcher with a subscription and good judgement.