Anthropic Restricts AI Model Rollout Over Cyberweapon Fears

Anthropic is limiting access to its most powerful AI model, Claude Mythos Preview, citing concerns that the technology could be weaponized by hackers faster than defenders can respond. The San Francisco-based AI safety company announced Project Glasswing on April 7, 2026, a restricted cybersecurity initiative that will provide Mythos capabilities to approximately 40 select companies for defensive security work only.

The decision marks a significant departure from the tech industry's typical rapid-deployment approach and underscores growing anxiety within Silicon Valley about AI's dual-use potential. Claude Mythos Preview excels at identifying software vulnerabilities and security flaws—capabilities that could equally empower cybercriminals, nation-state spies, or terrorist organizations if released broadly.

Amazon Web Services, Apple, Google, Microsoft, Nvidia, Cisco, JPMorgan Chase, CrowdStrike, and Palo Alto Networks are among the initial launch partners. These companies will use Mythos to find bugs in their own software systems and test whether specific hacking techniques work against their products. The Linux Foundation, Broadcom, and roughly 30 other firms building critical infrastructure will also participate.

The model has already demonstrated unprecedented capability in vulnerability discovery. According to Anthropic, Mythos identified thousands of previously unknown software flaws in recent weeks—a rate far exceeding human security researchers. In one notable case, the AI detected a 27-year-old bug in OpenBSD, a security-focused operating system that had remained undetected since 1999.

Anthropic executives briefed senior U.S. government officials across multiple agencies on Mythos's full offensive and defensive cyber capabilities before the announcement. Discussions included the Cybersecurity and Infrastructure Security Agency and the Center for AI Standards and Innovation. The company has made itself available to support government testing and evaluation of the technology.

"We did not feel comfortable releasing this generally," Logan Graham, who heads Anthropic's AI model defenses team, told CNN. "We think that there's a long way to go to have the appropriate safeguards."

The announcement follows a leak last month when Fortune discovered descriptions of Mythos in a publicly accessible data cache. The leaked blog post claimed the model was "far ahead" of competing AI systems in cyber capabilities and warned that it "presages an upcoming wave of models that can exploit vulnerabilities in ways that far outpace the efforts of defenders." Cybersecurity stocks fell sharply on the report before stabilizing.

Industry experts describe the situation as a fundamental shift in the cybersecurity landscape. AI agents can scan for vulnerabilities and potentially exploit them with speed and persistence impossible for human hackers. A single AI system could accomplish what previously required hundreds of skilled attackers working collaboratively.

"Cybersecurity is just going to be an area where this broad increase in capabilities has potential for risk, and thus we have to keep a really close eye on what's going on there," said Newton Cheng, Anthropic's Frontier Red Team cyber lead.

Project Glasswing derives its name from the glasswing butterfly, whose transparent wings metaphorically represent software vulnerabilities that remain "relatively invisible" until exposed. Anthropic has committed up to $100 million in usage credits for participating companies, though partners will pay for usage beyond that threshold.

The initiative comes weeks after Anthropic's high-profile clash with the Defense Department over AI safety concerns. The company, founded in 2021 by researchers who defected from OpenAI over safety disagreements, has positioned itself as more cautious about responsible AI deployment than competitors.

Dario Amodei, Anthropic's CEO, addressed the balance between risk and opportunity in a social media post announcing Project Glasswing: "The dangers of getting this wrong are obvious, but if we get it right, there is a real opportunity to create a fundamentally more secure internet and world than we had before the advent of AI-powered cyber capabilities."

Anthropic emphasized that Claude Mythos Preview is a general-purpose model not specifically trained for cybersecurity. Its advanced capabilities emerge from strong coding and reasoning skills developed during broader training. The company does not plan to make the model generally available but aims to learn how to eventually deploy Mythos-class models at scale with appropriate safeguards.

For investors and enterprises, the implications extend beyond immediate cybersecurity concerns. The restricted rollout signals that AI companies may face increasing pressure to implement capability gating—limiting access to powerful features based on use case and user verification. This could reshape go-to-market strategies for AI vendors and create competitive advantages for firms with early access to defensive tools.

The cybersecurity ETF market showed muted reaction to the announcement, with the iShares Cybersecurity ETF trading flat during intraday sessions. Analysts suggest that while short-term disruption is possible, long-term demand for AI-enhanced security tools will likely accelerate regardless of deployment restrictions.

As AI capabilities continue advancing, the gap between attackers and defenders remains the central challenge. Gadi Evron, founder of AI security firm Knostic, noted that defenders must adopt similar capabilities to maintain parity: "Unlike attackers, defenders don't yet have AI capabilities accelerating them to the same degree. However, the attack capabilities are available to attackers and defenders both, and defenders must use them if they're to keep up."

The coming months will test whether Project Glasswing's selective approach can meaningfully improve global software security before adversarial actors develop comparable capabilities independently.

Sources: CNBC, CNN Business, SecurityWeek, TechCrunch, IBM Think, Fortune This story is developing.