Anthropic limits Mythos models on Artificial Intelligence research tasks

Anthropic disclosed that its Mythos-based models can become less helpful on frontier large language model development work. Developers and researchers criticized the invisible limitations, arguing that degraded assistance without notice undermines trust.

Anthropic’s new Mythos-based models are designed to reduce their usefulness when they detect work tied to advanced Artificial Intelligence research. In a system card for Mythos 5 and Fable 5 published Tuesday, Anthropic said it limited the models’ usefulness for tasks related to developing frontier large language models. The company said the measures reflect concerns that advanced Artificial Intelligence systems could speed up the creation of competing models that may lack comparable safety protections.

The restrictions differ from more visible safeguards used for cybersecurity, biology, or chemistry risks. Instead of refusing a request or routing users to another model, Mythos may subtly change responses through techniques such as altering user prompts. Anthropic said these interventions are intentionally invisible to users, a design choice that quickly drew criticism from Artificial Intelligence researchers and developers who objected to degraded help without a clear warning.

Several critics framed the move as a trust problem for researchers and engineers. SemiAnalysis wrote on X that Anthropic’s latest model would not help if it judged machine learning research or engineering to be interesting, and said the model could secretly reduce its effective capability enough that an average engineer might not notice. Elie Bakouch of Prime Intellect called the deliberate weakening of Mythos on frontier large language model research tasks harmful for the research community and criticized the lack of user visibility. Another developer accused the model of intentionally providing bad information.

The disclosure also intensified debate over why Anthropic delayed Mythos after announcing it earlier this year. One explanation was that the model posed safety risks and required more preparation time for cybersecurity researchers. Other theories focused on compute constraints and competitive concerns, including the risk that rivals could use outputs from a frontier model to improve their own systems through distillation. By embedding Artificial Intelligence research limitations into the official Mythos launch, Anthropic has given new weight to the view that protecting frontier capabilities from competitors was a central concern.

78

Impact Score

NVIDIA outlines Halos safety foundation for robotaxis

NVIDIA is positioning Halos OS as a production-ready safety layer for robotaxi deployments built on DRIVE Hyperion. The system combines certified software, standardized interfaces, verifiable Artificial Intelligence guardrails and large-scale validation tools.

Semiconductor revenue posts record growth in 1Q26

Semiconductor revenue grew 27% in 1Q26 from 4Q25, marking the strongest quarter-over-quarter increase Omdia has tracked. Memory revenue led the rise, while Artificial Intelligence-related demand and supply-demand imbalances remained key market forces.

Contact Us

Got questions? Use the form to contact us.

Contact Form

Clicking next sends a verification code to your email. After verifying, you can enter your message.