Trump Lifts Critical Restrictions on Anthropic's AI Models, Igniting New Deployments

Anthropic has announced the restoration of public access to its highly capable Fable and Mythos frontier models, alongside the commercial launch of Claude Sonnet 5. This move concludes an eighteen-day operational pause that began on June 12, triggered by a U.S. government export control directive. The directive mandated a temporary suspension of Anthropic’s most advanced systems after Amazon cybersecurity researchers documented a method to bypass Fable 5’s safety controls, allowing the model to identify software vulnerabilities and generate exploitation code.

The initial restriction, enacted by government officials, underscored the significant regulatory pressures facing developers of frontier intelligence systems. The absence of real-time nationality verification systems necessitated a complete access blackout for all global users of Fable 5 and Mythos 5. Subsequent security evaluations revealed that the vulnerability identification behavior was not exclusive to Fable 5, with older, less capable architectures from various providers, including Claude Opus 4.8, GPT-5.5, and Kimi K2.7, exhibiting similar results.

To resolve the federal mandate, Anthropic engineers developed an updated automated safety classifier specifically targeting the bypass mechanism reported by Amazon. This new software layer operates with a wide safety margin, designed to identify and block ambiguous developer prompts that statistically indicate malicious intent. Internal validation data confirms that this classifier prevents the reported exploitation technique in over 99 percent of trials. When a prompt triggers this safety boundary, the platform automatically routes the workload to the older Opus 4.8 architecture to maintain service continuity. While enhancing safety, this expanded margin occasionally flags benign requests during routine application development and software debugging.

The U.S. government has since lifted the requirement for Anthropic to obtain an export license for its Mythos and Fable models. This decision, following weeks of discussions, allows Anthropic to resume global access. Secretary of Commerce Howard Lutnick confirmed that Anthropic has agreed to proactively detect and address security risks, collaborate with the U.S. government on protocols and standards for current and future models, and report any malicious activity. Cybersecurity experts had initially expressed skepticism about the restrictions, viewing them less as a security fix and more as potential political leverage by the Trump administration.

Amid these regulatory developments, Anthropic's immediate commercial focus is on the newly deployed Claude Sonnet 5. Engineering teams are actively transitioning autonomous agents to this model to optimize operational expenditure while maintaining high execution capacity. Performance data validates Sonnet 5's ability to execute multi-step plans, operate terminal environments, and navigate web browsers autonomously. Sonnet 5 demonstrates a SWE-bench Pro score of 63.2% and a Terminal-Bench 2.1 score of 80.4%, with introductory rates of $2.00 input and $10.00 output per million tokens through August 31, 2026, compared to its standard rates of $3.00 input and $15.00 output.

Real-world deployments showcase Sonnet 5's capabilities in live software development pipelines. Rakuten technology teams have deployed the architecture to process challenging production code pull requests, autonomously executing tests and verifying results before presenting completed code to human engineers for final approval. Software automation firm Zapier integrated Sonnet 5 into its core workflows for multi-part administrative tasks, such as updating Salesforce account tiers and transmitting launch announcements, tasks where prior models often stalled. Development tool provider Zed utilized Sonnet 5 to automate complex debugging procedures, where the system independently generated reproducing test scripts, applied fixes, and verified results without explicit prompts. Software engineering platform Factory reported that Sonnet 5 maintained logical grounding and execution consistency across corporate code repositories, outperforming previous generation software layers.

Quantitative safety audits indicate that Sonnet 5 achieves these autonomous capabilities without increasing security risks. Automated behavioral audits designed to detect deceptive tendencies and cooperation with unauthorized requests show a lower overall rate of non-compliant behavior compared to its predecessor, Sonnet 4.6. The architecture deliberately lacks advanced offensive cybersecurity capabilities, as Anthropic engineers omitted specialized cybersecurity datasets from its training. In public security assessments with Mozilla, Sonnet 5 failed to generate a single working exploit for known Firefox 147 browser core vulnerabilities, achieving a zero percent success rate. A 13.2 percent partial success rate was attributed to general logical reasoning gains rather than specific offensive training. Commercial versions of Sonnet 5 ship with default real-time safety classifiers equivalent to those used in the premier Opus 4.8 framework.

The regulatory friction surrounding Fable 5 prompted a formal partnership between Anthropic, Amazon, Microsoft, and Google to establish an objective industry framework for assessing model security breaches. This initiative addresses the current lack of shared metrics for classifying system bypass severity, which creates regulatory uncertainty. The proposed governance framework scores security breakdowns across four technical criteria: Capability gain, Breadth of capability gain, Ease of weaponization, and Discoverability. This matrix will help developers and cybersecurity professionals coordinate defensive responses, with automated mitigations deployed instantly for high-severity breaches. This effort complements a new HackerOne vulnerability research program and a dedicated corporate monitoring team providing 24-hour threat intelligence oversight.

Deployment strategies must adapt to this closer relationship between model builders and state regulatory bodies. Anthropic has formalized agreements under recent executive mandates to grant federal researchers early access to frontier architectures prior to public commercial release. These joint evaluation windows allow external security analysts to audit model capabilities alongside internal engineering teams, ensuring regulatory alignment before production. The Trump administration’s approach to AI policymaking, including the executive order on AI oversight, has been criticized for lacking clarity, particularly with similar restrictions imposed on OpenAI’s GPT-5.6 Sol, which is also accessible only to a select group of government-approved customers. Furthermore, the U.S. government faced pressure to ease restrictions on Anthropic to maintain global competitiveness against Asian AI companies like Fugu and Tulongfeng, which are developing models with Mythos-level capabilities.