Anthropic's Claude Mythos: how can security leaders prepare?

Anthropic’s recent release of the Claude Mythos Preview System Card sent the security industry into a tailspin. The Mythos update showed a significant improvement over previous versions of Anthropic’s LLM in autonomously identifying vulnerabilities and exploit chains across a number of software and operating systems. One key benchmark, CyberGym, showed an efficacy improvement from 66.6 to 83.1%. Results also showed it successfully exploited vulnerabilities in Firefox 147 at a rate of 72.4%, compared to the previous ~0.8% success rate.

For years, advanced exploit development has been an artisan craft that was performed by seasoned experts with deep expertise. Some of the most exquisite bugs may have taken weeks or months of effort, and even simple bugs can be days to work through for a determined, skilled developer. The ecosystem is shifting, and the barrier for entry has now plummeted, with weaponised exploit development dropping from weeks of expensive, purely human manual effort to a shorter timeline and just US$50 in AI costs for one example (CVE-2026-4747).

The most worrying behaviours Anthropic found weren’t intentional. As the AI improved coding and reasoning skills and began acting independently, it exposed a glimpse of what AI can become without proper rules or safeguards in place. Thankfully, the Mythos Preview release had already benefited from alignment optimisation work to improve the safety guardrails. But Anthropic’s diligence will only last for so long, with copycat tools imminent.

While many have been warning of the need to balance the opportunities and dangers of AI, it feels like we’re now standing at an inflexion point. One where the creators of AI tools need to be incredibly careful that the technology doesn’t fall into the wrong hands, while security teams need to wake up to the reality that it’s only a matter of time until they do. In fact, breaches are already happening, with reports that a private Discord group gained unauthorised access to the Claude Mythos preview through a third‑party vendor environment.

Security professionals have long been concerned about the growing capabilities of frontier models and the capabilities of security teams to keep up. These are no longer theoretical advances, though the same capabilities that may be weaponised by an adversary can be used to defend at the same pace. Importantly, even the most powerful new models still struggled to successfully compromise fully patched systems that implemented best-practice hardening like the Essential 8 controls: patching, multi-factor authentication, restricted administrative control, and reduced macro exposure. These foundational security tools all stand firmly in the way of an AI-powered adversary.

Exploit development is no longer a professional’s game

What stands out with Mythos is not just the improvement on previous models, but the scale at which it improved against the same class of targets. This suggests exploit development is no longer a high-skill, high-effort bottleneck. As with the rise of ‘vibe coding’, we’re now facing a world where exploit development has gone from a master craft to something just about anyone with an internet connection could try their hand at if given the right tools.

The timing matters too. The gap between frontier models and open-weight models has compressed from more than a year to a matter of weeks or months, which means this level of capability is poised to spread rapidly, likely without the same safety guardrails. Anthropic put a lot of work into making Mythos safer, but that same safety doesn’t automatically come when others start to build on the blocks Anthropic has created.

Meanwhile, disclosure-to-exploitation timelines are expected to collapse from days to hours. The previous, often-siloed workflows that involved scan, review, patch, repeat cycles on a monthly cadence are outdated. In response to the release of the Mythos Preview, the Australian Signals Directorate is now encouraging organisations to patch urgently and efficiently, something that is not possible without complete visibility and autonomous scanning and remediation. Continuous threat exposure management and real-time verification of patch efficacy are now bare minimum requirements.

Visibility: the first thing security leaders should be prioritising

Preventing predictable problems is no longer enough when AI is now lightning fast and unpredictable. To meet the moment, security teams need to see rather than predict. The exposure-to-remediation loop needs to be closed before bad actors can leverage AI to take advantage of a new zero-day vulnerability. This takes real-time visibility.

Organisations, more than ever, need wall-to-wall visibility. This means real-time awareness of all endpoints, including on-premise, cloud, remote and OT. You can’t beat AI if you can’t see every potential point of vulnerability. A lack of visibility undercuts basic processes like patching, and it severely impairs incident response.

Secondly, organisations should keep a continuous Software Bill of Materials (SBOM)-grade inventory that keeps pace with AI-accelerated disclosure cycles. The Cloud Security Alliance highlighted SBOMs as a high priority for the next 90 days in a post-mythos readiness world. The message is clear that organisations must implement the infrastructure and tooling today that enable them to respond to future vulnerabilities, no matter where they pop up tomorrow. If a new vulnerability drops, security teams no longer have time to waste figuring out where that component sits, if they’re affected and whether it needs to be patched or not. With Mythos-type AI tools, bad actors will have exploited it before you can react. Organisations need to be aiming for sniper precision patching, or else leave themselves incredibly vulnerable.

Anthropic’s Claude Mythos has changed the game. AI finds vulnerabilities in hours and exploits cost less than a tank of petrol to build. In July 2026, the results and patches associated with Project Glasswing’s findings will make their way into patch workflows and backlogs across the largest providers of software and operating systems in the world. Preparing for this surge and the modern cadence of vulnerability discovery to disclosure and then patch deployment is the new reality. The threat is real, and it is here now. Security teams cannot continue to run remediation cycles on manual workflows. The only way to win it is to operate at machine speed through wall-to-wall visibility and real-time, autonomous patching.

Image credit: iStock.com/Olemedia