Anthropic Resumes Access to Frontier AI Models After Export Control Review, Launches Claude Sonnet 5
Anthropic has relaunched its Fable and Mythos frontier AI models, restoring full access following a federal export control review. This decision marks the end of an eighteen-day operational pause that began on June 12, when a U.S. government directive compelled the temporary suspension of Anthropic’s most advanced AI systems.
The restriction was put in place after researchers at Amazon documented a method to bypass the safety controls of Fable 5, enabling the model to identify software vulnerabilities and generate exploitation code. Anthropic has since developed and deployed an updated automated classifier to address this vulnerability, paving the way for a comprehensive commercial rollout across its platform, cloud infrastructure, and partner networks.
The temporary suspension of Fable 5 and Mythos 5 underscored the significant regulatory scrutiny now facing frontier AI systems. The initial export control mandate necessitated a complete, global access blackout due to the absence of real-time nationality verification systems capable of distinguishing between authorized and unauthorized users.
Crucially, security evaluations conducted during the shutdown revealed that the vulnerability identification behavior was not unique to Fable 5. Older, less capable architectures from multiple providers, including Claude Opus 4.8, GPT-5.5, and Kimi K2.7, demonstrated the same output when subjected to similar prompts. This finding highlights a systemic challenge in AI safety across the industry rather than an isolated flaw.
To satisfy the federal directive, Anthropic engineers trained a specialized automated safety classifier specifically targeting the bypass mechanism identified by Amazon. This software layer is designed with a wide safety margin, effectively identifying and blocking ambiguous developer prompts that exhibit a statistical probability of malicious intent. Internal validation data suggests this updated classifier prevents the reported exploitation technique in over 99% of test scenarios.
When a developer’s prompt triggers this safety boundary, the platform automatically reroutes the workload to the older, more robust Opus 4.8 architecture to ensure operational continuity. However, this expanded safety margin introduces a trade-off for development teams. The automated system may flag benign requests more frequently during routine application development and software debugging, potentially slowing down iterative processes.
**Focus Shifts to Claude Sonnet 5 for Commercial Applications**
While the frontier models navigate stringent governmental oversight, the immediate commercial focus for Anthropic is on the newly deployed Claude Sonnet 5. Engineering teams are actively migrating autonomous agents to this model, aiming to reduce operational expenditures while maintaining high execution capabilities. Performance data indicates that Sonnet 5 can execute multi-step plans, operate within terminal environments, and navigate web browsers autonomously.
| Model | SWE-bench Pro | Terminal-Bench 2.1 | Base Input Cost* ($ per million tokens) | Base Output Cost* ($ per million tokens) |
| :———– | :———— | :—————– | :———————————— | :————————————- |
| Sonnet 5 | 63.2% | 80.4% | $3.00 | $15.00 |
| Sonnet 4.6 | 58.1% | 67.0% | $3.00 | $15.00 |
| Opus 4.8 | 69.2% | 82.7% | $5.00 | $25.00 |
* *Sonnet 5 currently offers introductory rates of $2.00 for input and $10.00 for output per million tokens through August 31, 2026.*
Real-world deployments illustrate how organizations are integrating this architecture into their live software development pipelines. At Rakuten, technology teams have successfully utilized the architecture to process challenging production code pull requests, independently executing tests and verifying results before human engineers provide final structural approval.
Software automation firm Zapier has integrated Sonnet 5 into its core product workflows to streamline complex administrative tasks. In one documented deployment, engineers tasked the model with updating Salesforce account tiers and subsequently generating and disseminating launch announcements to enterprise contacts. Previous model architectures often faltered midway through such multi-stage operations, whereas the current system has demonstrated the ability to execute entire sequences end-to-end without human intervention.
Development tool provider Zed has leveraged Sonnet 5 to automate intricate debugging procedures. During internal trials, engineering teams directed the model to investigate an active software bug. Operating without explicit step-by-step instructions, the system autonomously generated a script to reproduce the bug, applied the necessary code fix, and then verified that the bug reappeared in the absence of the patch. This entire diagnostic and remediation cycle was completed within a single processing pass.
Furthermore, the software engineering platform Factory has implemented the architecture to manage sustained coding tasks within complex codebase environments. Technical teams have reported that Sonnet 5 maintains logical grounding and execution consistency across extensive corporate code repositories, outperforming previous generation software layers by completing tasks that previously timed out or failed to resolve.
**Quantitative Safety Audits and Limitations**
Anthropic’s system card data indicates that these advanced autonomous capabilities are achieved without a commensurate increase in security risks. Automated behavioral audits, designed to detect deceptive tendencies and cooperation with unauthorized requests, show that Sonnet 5 exhibits a lower overall rate of non-compliant behavior compared to its predecessor, Sonnet 4.6.
It is important to note that the architecture does not possess advanced offensive cybersecurity capabilities. Anthropic engineers deliberately excluded specialized cybersecurity datasets from the training protocol, limiting the system’s utility to routine, defensive technical tasks. In public security assessments conducted in partnership with Mozilla, researchers tested the model’s capacity to generate functional exploits for known vulnerabilities within the Firefox 147 browser core. The model failed to generate any working exploits across all evaluation windows, achieving a zero percent success rate. It did, however, record a 13.2% partial success rate, a minor increase attributed by engineers to general gains in logical reasoning rather than domain-specific offensive training. As a precautionary measure, commercial versions are deployed with default real-time safety classifiers equivalent to those used in the premier Opus 4.8 framework.
**Industry-Wide Collaboration on AI Security Frameworks**
The regulatory friction surrounding Fable 5 has catalyzed a formal partnership between Anthropic, Amazon, Microsoft, and Google. The objective is to establish an objective industry framework for assessing AI model security breaches. Currently, AI providers lack a standardized metric to classify the severity of system bypasses, leading to regulatory uncertainty when researchers uncover new prompting vulnerabilities.
The proposed governance framework will score security breakdowns across four specific technical criteria:
* **Capability Gain:** Measures the extent to which an exploit enhances user capabilities beyond standard, widely available software utilities.
* **Breadth of Capability Gain:** Quantifies the number of distinct offensive operations a single exploit can unlock.
* **Ease of Weaponization:** Tracks the human engineering effort and specialized prompting required to elicit a harmful output.
* **Discoverability:** Determines the accessibility of the exploit technique within public research circles.
This matrix will serve as a vital tool for developers and cybersecurity professionals to coordinate defensive responses. For high-severity breaches, such as exploits with the immediate potential to disrupt financial systems or critical infrastructure, providers will be empowered to deploy automated mitigations instantly. This initiative is complemented by a newly established HackerOne vulnerability research program and a dedicated corporate monitoring team providing 24-hour oversight of threat intelligence channels.
Deployment strategies will undoubtedly need to adapt to this closer collaboration between AI 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 their public commercial release. These joint evaluation windows will enable external security analysts to audit model capabilities alongside internal engineering teams, ensuring regulatory alignment before code enters production environments.
Original article, Author: Samuel Thompson. If you wish to reprint this article, please indicate the source:https://aicnbc.com/23333.html
