AI cybersecurity is now a proper aggressive entrance between OpenAI and Anthropic, with OpenAI finalizing a complicated safety product for a restricted accomplice launch and Anthropic operating a tightly managed effort referred to as Challenge Glasswing geared toward discovering crucial software program vulnerabilities earlier than attackers do.
Abstract
- OpenAI is finalizing an AI cybersecurity product for launch first to a restricted set of companions.
- Anthropic’s Challenge Glasswing is a managed initiative centered on looking crucial software program vulnerabilities proactively.
- Each efforts elevate basic questions on who controls AI offense and protection instruments and who’s accountable when issues go improper.
Synthetic intelligence has moved from a software that helps defenders perceive threats to 1 that may independently discover and exploit vulnerabilities. OpenAI and Anthropic at the moment are constructing immediately into that area, with implications for governments, enterprises, and the thousands and thousands of software program methods that underpin international monetary infrastructure.
OpenAI is finalizing an AI cybersecurity product with superior capabilities and plans to launch it initially to a restricted accomplice group, in line with Tech Startups. Anthropic is operating a parallel effort internally referred to as Challenge Glasswing, a tightly managed initiative designed to seek out crucial software program vulnerabilities earlier than malicious actors discover them first.
The twin bulletins mark a shift in how the 2 main AI labs are positioning themselves. Each are transferring from general-purpose AI into security-specific merchandise with direct offensive and defensive functionality. The query is now not what AI can do in cybersecurity. It’s who controls it and who’s accountable when it goes improper.
What Anthropic’s Monitor File Exhibits
Anthropic has already demonstrated the dimensions of what AI safety instruments can obtain. As crypto.information reported, the corporate restricted entry to its Claude Mythos Preview mannequin after early testing discovered it might uncover 1000’s of crucial vulnerabilities throughout extensively used software program environments, together with a 27-year-old bug in OpenBSD and a 16-year-old distant execution flaw in FreeBSD. Anthropic stated: “Given the speed of AI progress, it won’t be lengthy earlier than such capabilities proliferate, doubtlessly past actors who’re dedicated to deploying them safely.”
Trade knowledge cited by Anthropic exhibits a 72% year-on-year enhance in AI-powered cyberattacks, with 87% of world organizations reporting publicity to AI-enabled incidents in 2025. Challenge Glasswing is being positioned as Anthropic’s managed effort to remain forward of that curve.
The Threat of Twin-Use AI Safety Instruments
The deeper situation for regulators and the trade is that the identical AI software that finds a vulnerability defensively can discover it offensively. As crypto.information famous, a joint examine by Anthropic and MATS Fellows discovered that Claude Sonnet and GPT-5 might produce simulated exploits in opposition to Ethereum sensible contracts value $4.6 million in testing, and uncovered two novel zero-day vulnerabilities in almost 3,000 lately deployed contracts.
That dual-use actuality makes the managed rollout methods each firms are pursuing important. However the query of whether or not restricted entry is sufficient to stop proliferation is one neither lab has totally answered.
