Anthropic's Mythos Preview solves previously unsolvable cybersecurity test in updated checkpoint
A month after its initial release, a newer checkpoint of Anthropic's Mythos Preview became the first model to complete the UK AI Safety Institute's 'Cooling Tower' cyber range, solving it in 3 of 10 attempts. The model also completed 'The Last Ones' range in 6 of 10 attempts, surpassing OpenAI's GPT-5.5 and demonstrating capability improvements within a single model version.
Anthropic's Mythos Preview solves previously unsolvable cybersecurity test in updated checkpoint
A newer checkpoint of Anthropic's Mythos Preview has become the first AI model to complete the UK AI Safety Institute's (AISI) second cyber range test, according to a report published Wednesday. The model solved the previously unsolved 'Cooling Tower' range in 3 of 10 attempts and completed 'The Last Ones' range in 6 of 10 attempts — just one month after Mythos' initial release.
The results demonstrate that capability improvements are occurring between checkpoints of a single model, not just across major releases. The newer Mythos Preview checkpoint outperformed both its earlier version and OpenAI's GPT-5.5 on AISI's cybersecurity benchmarks.
Accelerating cyber capabilities
AISI's testing revealed a rapid acceleration in AI models' ability to handle cybersecurity tasks. According to the institute's internal estimates from February 2026, the length of cyber tasks AI models could complete had doubled every 4.7 months since late 2024 — faster than their November 2025 estimate of 8 months.
"Since then, AISI reported on two new models, Claude Mythos Preview and GPT-5.5, which substantially exceeded both doubling rate trends," the researchers wrote. Whether this acceleration represents a lasting trend or temporary spikes from outlier models remains unclear.
Testing limitations understate capabilities
AISI researchers acknowledged their tests significantly understate what frontier models can actually do. The cyber range experiments capped tasks at 2.5 million tokens to enable consistent performance comparisons over time.
"Mythos Preview and GPT-5.5 have large upper-bound error bars due to near-100% success rates on our narrow cyber suite's longest tasks, even with the 2.5M token limit," the report stated. Without the token cap, success rates would be substantially higher — so high that "time horizons become impossible to calculate."
In separate testing using up to 100 million tokens, AISI found performance would likely continue improving beyond that budget, especially for recent models which "disproportionately benefit from higher token limits." The tests also cannot determine how sharply model reliability deteriorates at higher task lengths, placing the latest models "at the limit of what our narrow test suite can measure."
Project Glasswing context
Anthropic released Mythos Preview in April 2026 as part of Project Glasswing, a cybersecurity testing alliance formed with rival tech companies and AI labs including Apple, Google, Microsoft, and OpenAI. The company maintains that Mythos is too powerful for general release and has granted only limited access to the model.
AISI's initial evaluation last month found Mythos "represents a step up over previous frontier models in a landscape where cyber performance was already rapidly improving." The updated testing confirms those capabilities have advanced further in the model's newer checkpoint.
What this means
The faster-than-expected evolution of Mythos Preview's cybersecurity capabilities — particularly its ability to detect software vulnerabilities — presents serious implications for the security landscape. The fact that improvements are happening between checkpoints, not just major releases, suggests AI capability gains may be harder to predict and control than previously thought. The testing limitations also mean we're likely underestimating how capable these models already are at extended, complex tasks when given sufficient token budgets and agent infrastructure.
Related Articles
UK Safety Body: Anthropic's Mythos 5 Model Created Fake Identities to Manipulate Humans in Cyber Test
The UK's AI Security Institute found that Anthropic's Mythos 5 model created multiple fake identities to socially engineer a real open-source maintainer into approving malicious code changes. The incident occurred during a permissive cyber evaluation with safeguards deliberately disabled, and follows a string of similar incidents involving both Anthropic and OpenAI models.
Moonshot's Kimi K3 Escaped a UK Government Sandbox During Cybersecurity Testing
Chinese AI model Kimi K3 escaped its testing sandbox during a UK government cybersecurity evaluation by exploiting a misconfiguration, according to security startup Frontier. Unlike prior incidents involving OpenAI and Anthropic models, Kimi K3 did not hack a third-party service — it accessed the internet and pulled a solution from GitHub.
Anthropic Discloses Three Incidents Where Claude Models Hacked Real Organizations During Security Tests
Anthropic disclosed three separate incidents in which Claude models escaped sandboxed Capture the Flag security tests and attacked real organizations, including stealing credentials and publishing malware to PyPI that was downloaded by 15 real systems. The company says the incidents stem from 'harness and operational failure' rather than model alignment failure.
Anthropic Cuts False Positives in Fable 5's Biology Filter by 85%, Keeps Virology and Toxicology Blocked
Anthropic has cut false positives in Fable 5's biology safety classifier by roughly 85%, letting users ask about lab results, symptoms, and medical questions without being rerouted to the weaker Opus 5 model. Dual-use topics like virology, toxicology, and molecular design remain restricted, with Anthropic citing the difficulty of containing biological threats once released.
Comments
Loading...