UK tests show Anthropic's Claude Mythos succeeds in autonomous cyberattacks
The UK's AI Security Institute tested Anthropic's Claude Mythos model and found it capable of autonomously executing multi-stage cyberattacks, succeeding in 73% of expert-level penetration tests and completing a full 32-stage enterprise network attack in some test runs. The agency raised concerns about the model's potential to compromise inadequately defended systems and recommended organizations strengthen basic cybersecurity practices.",antml:parameter> <parameter name="event_date">2026-04-13
What happened
The UK's AI Security Institute (AISI) tested Anthropic's Claude Mythos model on April 13, 2026, and found it could autonomously conduct multi-stage cyberattacks.
In a capture-the-flag (CTF) exercise, Mythos achieved a 73% success rate on expert-level tasks that no model had previously solved before April 2025.
In a 32-stage simulated enterprise network attack expected to take a human expert approximately 20 hours, Mythos completed the full attack sequence in 3 out of 10 attempts—the first model to do so—and reached an average of 22 stages across all attempts.
Reported impact
AISI assessed that Mythos has the capability to autonomously attack small-to-medium enterprise systems with weak defenses that have already granted network access.
Experts cited the need for international rules to prevent misuse of the model's capabilities.
Index analysis
AISI noted that its evaluation lacked active defense personnel, detection tools, or penalties for attacks, so results in fully defended real-world environments remain uncertain.
AISI recommended organizations implement basic security measures including regular security updates, robust access controls, proper security configuration, and comprehensive logging.
Governance controls
AISI indicated plans to evolve evaluation methods by transitioning to assessments in more realistic defensive environments with real-time detection and incident response capabilities.
AISI emphasized that AI's cyber capabilities could serve both attack and defense functions, expressing expectations for defensive applications.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.