Anthropic has taken the weird step of withholding its most succesful AI mannequin from the general public, not as a result of it underperforms, however as a result of it really works too properly. Claude Mythos Preview, announced Tuesday as a part of a brand new business initiative known as Venture Glasswing, has demonstrated cybersecurity capabilities so superior that the corporate has concluded a basic launch would pose unacceptable dangers.
The mannequin has already recognized hundreds of beforehand unknown zero-day vulnerabilities — together with essential flaws in each main working system and each main net browser — a lot of which have survived many years of human overview and hundreds of thousands of automated safety exams. One vulnerability it found in OpenBSD, extensively considered some of the security-hardened working techniques on the planet, had gone undetected for 27 years. One other, within the ubiquitous video-processing library FFmpeg, had been missed regardless of automated testing instruments hitting the related line of code 5 million instances.
Reasonably than making Mythos out there by means of its API, Anthropic has assembled a coalition of 12 launch companions — together with Amazon Net Providers, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Basis, Microsoft, NVIDIA, and Palo Alto Networks — to deploy the mannequin completely for defensive safety work. A further 40 organisations that construct or keep essential software program infrastructure can even obtain entry. Anthropic is committing as much as $100 million in utilization credit and $Four million in direct donations to open-source safety organisations to assist the trouble.
A Mannequin That Thinks Like a Safety Researcher
What separates Mythos from its predecessors isn’t just the amount of vulnerabilities it may well discover, however its capability to function with minimal human steering. In line with Anthropic, the mannequin recognized almost all of the disclosed vulnerabilities — and developed working exploits for a lot of of them — completely autonomously, with out human steering. In a single case, it independently found and chained collectively a number of Linux kernel vulnerabilities to escalate from bizarre person entry to full machine management, a feat that sometimes requires in depth guide experience.
Logan Graham, who leads Anthropic’s frontier crimson group, instructed Axios that Mythos Preview is “extraordinarily autonomous” and possesses reasoning capabilities corresponding to a sophisticated human safety researcher. The place its predecessor, Claude Opus 4.6, discovered roughly 500 zero-days in open-source software program, Mythos Preview’s output runs into the tens of hundreds.
On the CyberGym vulnerability copy benchmark, Mythos scored 83.1% in comparison with Opus 4.6’s 66.6%. The efficiency hole extends throughout coding and reasoning duties extra broadly: on SWE-bench Verified, a measure of real-world software program engineering functionality, Mythos scored 93.9% towards Opus 4.6’s 80.8%.
“AI capabilities have crossed a threshold that essentially adjustments the urgency required to guard essential infrastructure from cyber threats, and there’s no going again,” stated Anthony Grieco, SVP and Chief Safety and Belief Officer at Cisco. “The previous methods of hardening techniques are not adequate.”
The Twin-Use Dilemma
The capabilities that make Mythos precious to defenders are exactly what make it harmful within the flawed fingers. This dual-use stress sits on the coronary heart of the Venture Glasswing announcement and represents some of the concrete examples but of the cybersecurity arms race that OpenAI’s recently published industrial policy paper warned about.
In that document, launched the day earlier than Anthropic’s announcement, OpenAI recognized AI-enabled cyberattacks and organic threats as the 2 most fast near-term dangers from superior AI. Sam Altman instructed Axios that high tech, enterprise and authorities officers worry that soon-to-be-released fashions may allow a world-shaking cyberattack this 12 months. The Mythos announcement lends appreciable weight to that warning.
The timing is just not coincidental. The broader AI business has been converging on the view that frontier fashions now pose real cybersecurity dangers at a scale that calls for coordinated motion. OpenAI warned in December that its personal upcoming fashions posed a “excessive” cybersecurity danger. CNN reported that specialists see the emergence of AI brokers — autonomous techniques able to scanning for and exploiting vulnerabilities far quicker than human hackers — as a step-change within the risk panorama.
“If we’re crossing the Rubicon the place you may functionally automate these capabilities and make them very low-cost as properly, then we’re in a completely new world,” Graham instructed the Washington Examiner.
Actual-world proof already helps this concern. Anthropic has beforehand disclosed {that a} Chinese language state-sponsored hacking group exploited Claude’s agentic capabilities to infiltrate roughly 30 organisations, together with tech corporations, monetary establishments and authorities companies. In February, a hacker used Claude in a collection of assaults towards Mexican authorities companies, stealing delicate tax and voter info.
The Glasswing Technique: Defence Earlier than Offence
Anthropic’s method — limiting the mannequin to vetted companions fairly than releasing it publicly — represents a notable departure from the final development in AI towards broad availability. The corporate is explicitly betting that giving defenders a head begin will create a sturdy benefit earlier than related capabilities proliferate by means of competing fashions.
“The window between a vulnerability being found and being exploited by an adversary has collapsed — what as soon as took months now occurs in minutes with AI,” stated Elia Zaitsev, CTO of CrowdStrike. “That’s not a cause to decelerate; it’s a cause to maneuver collectively, quicker.”
Jim Zemlin, CEO of the Linux Basis, emphasised the implications for open-source software program, which underpins the overwhelming majority of contemporary digital infrastructure. Open-source maintainers — usually volunteers working with out devoted safety groups — have traditionally been liable for software program that runs the world’s servers, banking techniques and communications networks. Venture Glasswing, Zemlin stated, provides a path to creating AI-augmented safety accessible to maintainers who beforehand couldn’t afford it.
Microsoft, which is concurrently a companion in Venture Glasswing and a significant investor in OpenAI, reported that when examined towards its CTI-REALM safety benchmark, Mythos Preview confirmed substantial enhancements over earlier fashions. Igor Tsyganskiy, EVP of Cybersecurity and Microsoft Analysis, stated the initiative permits the corporate to “establish and mitigate danger early.”
Questions of Belief and Timing
The announcement doesn’t come with out problems. Anthropic’s personal safety monitor report has just lately been examined: final month, the corporate unintentionally uncovered almost 2,000 supply code recordsdata and over half 1,000,000 traces of code by means of a mistake in its Claude Code software program package deal. The existence of Mythos itself was first revealed by means of a leak from an unsecured, publicly searchable knowledge cache — an ironic safety lapse for an organization positioning itself as a cybersecurity chief.
Anthropic has additionally been navigating a high-stakes dispute with the U.S. Division of Defence over the army’s use of Claude in real-world operations, including geopolitical complexity to its authorities engagement round Mythos. The corporate says it has been briefing CISA, the Commerce Division and different companies on the mannequin’s capabilities, framing the initiative in nationwide safety phrases.
The broader query — raised by OpenAI’s industrial coverage paper and now given tangible kind by Mythos — is whether or not the establishments, rules and coordination mechanisms wanted to handle AI-driven cybersecurity danger can maintain tempo with the expertise itself. Anthropic has dedicated to publishing a public report on classes realized inside 90 days. Companions will share info and finest practices. Sensible suggestions for the way vulnerability disclosure, patching, and software program growth practices ought to evolve within the AI period are anticipated to observe.
However as Graham acknowledged, Mythos is just not the tip of the story. Behind it sits the following OpenAI mannequin, the following Google Gemini, and some months behind them, open-source Chinese language fashions that might replicate related capabilities with out the guardrails or coalition-based deployment that Venture Glasswing represents.
“Nobody organisation can clear up these cybersecurity issues alone,” Anthropic wrote. “The work of defending the world’s cyber infrastructure may take years; frontier AI capabilities are more likely to advance considerably over simply the following few months.”
The race between AI-powered assault and AI-powered defence is now absolutely underway. Venture Glasswing is Anthropic’s opening bid that the defenders can win. Whether or not it proves adequate will rely on how shortly the remainder of the business, and the world’s governments, observe.
Jason Jones Jason Jones Read More
