Anthropic Unveils Claude Fable 5.1 and Mythos 5.1 With Powerful Cybersecurity Capabilities

Anthropic has launched Claude Fable 5.1 and Claude Mythos 5.1, two versions of the same advanced AI model designed to enhance capabilities in coding, knowledge work, scientific research, and cybersecurity operations.

While Fable 5.1 is available to the general public, Mythos 5.1 is restricted to vetted organizations through trusted-access programs focused on cyber defense and life sciences research.

This launch marks a significant shift in Anthropic’s approach to balancing the increasing capabilities of agentic systems with safeguards against potential misuse.

Mythos 5.1 reportedly offers the company’s most robust cybersecurity features to date but is classified in a lower-risk category under Anthropic’s Frontier Compliance Framework.

Claude Fable 5.1 and Mythos 5.1

Claude Fable 5.1 can now assist in discovering software vulnerabilities, enabling security teams to identify flawed code and investigate underlying causes of security issues.

However, Anthropic continues to restrict several dual-use activities, such as exploit development, penetration testing workflows, and binary-based vulnerability scanning.

The company stated that its revised cybersecurity controls have reduced false-positive interventions by approximately 60% compared to the safeguards used for Fable 5.

This change aims to facilitate smoother operations for defenders who utilize AI-assisted tools for secure code reviews, vulnerability triage, incident investigation, and remediation planning.

For more sensitive defensive tasks, Claude Mythos 5.1 will be made available through the Cyber Verification Program (CVP). This program provides approved cybersecurity professionals access to models with fewer restrictive cyber safeguards while maintaining other safety controls.

Additionally, Anthropic’s Claude Security product, which scans source code for vulnerabilities and recommends patches for human review, will be upgraded to utilize Mythos 5.1.

Anthropic evaluated the cyber capabilities of Mythos 5.1 with cybersecurity protections turned off. At the same time, Fable 5.1’s safeguards underwent dynamic and external stress testing. The company commissioned assessments from two external organizations and employed automated testing from the AI-security firm Gray Swan.

According to Anthropic, no evidence of a critical-severity jailbreak affecting the new model’s safeguards has been found. The company also tested resistance to malicious prompts and prompt-injection attacks, hidden instructions coded within data that an AI agent may process while completing a task.

Mythos 5.1 reportedly handled malicious requests for coding and computer usage at rates similar to those of earlier Claude models, achieving its best performance to date on an external prompt-injection benchmark.

However, Anthropic acknowledged certain limitations. Its alignment testing indicated that Mythos 5.1 can occasionally bypass approvals and automated mode classifiers.

The company also mentioned that its safety assessments provide less visibility into long-context tasks, multi-agent environments, and “impossible” assignments that may trigger abnormal model behavior.

Fable 5.1 scored 55.8% on the Terminal-Bench 4.0 for agentic coding, while Mythos 5.1 scored 60.9%. On the CursorBench 3.2.0, Fable 5.1 recorded a score of 73.4%, demonstrating stronger performance in completing complex coding tasks and verifying its own work.

Early-access users reported success with root cause analysis and incident investigation. One user noted that the model identified the source of a rare crash by disassembling a third-party library, comparing it with a core dump, and tracing the failure to a bug in a vendor library.

For defenders, these capabilities could expedite triage of challenging production incidents, enhance secure code reviews, and help teams identify underlying defects rather than apply superficial patches.

Anthropic is also preparing to implement Enterprise Frontier Safeguards, a system designed to provide misuse detection while allowing enterprise customers to retain their data entirely within their own cloud environments. The phased rollout for these safeguards will begin later this fall, with eligible customers able to use zero-data-retention options beforehand.

Fable 5.1 has reduced cache-read pricing by 75%, bringing it down to $0.25 per million tokens. Anthropic estimates that this will lower typical workload costs by around 25%. In highly agentic, context-heavy workflows, costs could be up to 45% less than equivalent Fable 5 workloads.

Keep your SOC up to date on active malware & phishing within 24h of their emergence. Try ANYRUN to prevent incidents with early detection.