AI Cyber Threats: How Claude Mythos and GPT-5.5 Are Revolutionizing Cybersecurity in 2026

Translation result.

Analysts warn that the cyber-attack capabilities of the latest AI models are advancing far faster than expected. In particular, Anthropic’s security-focused model Claude Mythos Preview and OpenAI’s GPT-5.5-Cyber have posted performance gains that outpace prior trends, heightening concerns about AI-driven cyber threats.

On the 13th (local time), the U.K. AI Safety Institute (AISI) said in a report that the complexity of cyber tasks these models can perform autonomously is rising sharply on a timescale measured in months.

AISI estimated that since late last year frontier AI models’ ability to carry out cyber tasks had been doubling roughly every 4.7 months. The institute says the newest models have exceeded even that rate of improvement.

Notably, the models evaluated were not the initial Mythos release Anthropic published in April, but a newer version of Mythos.

The institute cautioned that it remains unclear whether this represents a one-time leap or the start of a faster development phase.

AISI measures a model’s “cyber time horizon” by comparing how long a human security expert would take to complete the same task. The metric indicates how lengthy and complex a hacking operation an AI can autonomously execute.

For example, the latest models reached a level where they can carry out a cyber-attack that would take a human expert about 16 minutes with an 80% success probability. Researchers capped tests at 2.5 million tokens per task for comparison, but noted that using more tokens and advanced agent architectures could significantly boost performance.

The newer Mythos model became the first to pass both of AISI’s high-difficulty cyber-attack scenarios. These scenarios assess an AI’s ability to run a complex campaign that, after initial intrusion, aims to take control of an entire corporate network—requiring sustained planning and execution.

In testing, Mythos succeeded in The Last Ones scenario 6 out of 10 times and achieved the previously unsolved Cooling Tower attack 3 out of 10 times. By contrast, GPT-5.5-Cyber succeeded in The Last Ones only 3 out of 10 attempts.

AISI judged that both models were capable of attacks that, in a real environment, could amount to a full network takeover. That marks a significant advance compared with the earlier Mythos version, which had been limited to the advanced persistence stage.

Anthropic has not made Mythos publicly available because of these risks. The company is providing the model only to select firms and government agencies for software-vulnerability detection and security hardening. OpenAI has similarly distributed GPT-5.5-Cyber to a small number of firms for testing.

Crucially, AI models are beginning not only to reproduce known attacks but to discover new vulnerabilities. The Mozilla Foundation reported that early Mythos versions found more than 100 high-risk flaws in the Firefox browser—roughly the number the industry typically uncovers in two months.

The security sector has coined the term “Bugmageddon” to describe this trend and warned of its implications. If AI discovers vulnerabilities much faster than human researchers, patching and defense systems may struggle to keep pace.

Market moves reflect that tension. So far this year, shares of major cybersecurity firms such as Palo Alto Networks and CrowdStrike have risen roughly 20% on expectations that AI-driven threats will spur corporate security spending.

Researchers acknowledged limits in the current evaluation framework: the number of test tasks was limited, and some human-expert time estimates are provisional. They also noted that results could differ in real-world environments where active defenses exist.

Still, AISI stressed that frontier AI’s autonomous cyber capabilities are evolving on a monthly, not multi-year, timescale. The institute said this is a critical moment to build security resilience, since AI can empower both attackers and defenders.

Reporter Chan Park cpark@aitimes.com