{"id":15525,"date":"2026-07-01T12:41:38","date_gmt":"2026-07-01T09:41:38","guid":{"rendered":"https:\/\/harici.com.tr\/en\/?p=15525"},"modified":"2026-07-01T12:42:11","modified_gmt":"2026-07-01T09:42:11","slug":"us-lifts-export-controls-on-anthropics-claude-5-models-after-security-upgrade","status":"publish","type":"post","link":"https:\/\/harici.com.tr\/en\/us-lifts-export-controls-on-anthropics-claude-5-models-after-security-upgrade\/","title":{"rendered":"US lifts export controls on Anthropic\u2019s Claude 5 models after security upgrade"},"content":{"rendered":"

Anthropic has announced<\/a><\/span> that the US government has lifted export controls on its newest artificial intelligence models, Claude Fable 5 and Claude Mythos 5, paving the way for the high-profile systems to return to active use.<\/p>\n

The AI research and development company had temporarily suspended access for all users following the June 12 imposition of the restrictions, which blocked foreign nationals both inside and outside the US from accessing the models. Anthropic stated that the complete suspension was necessary because it was unable to perform real-time nationality verification. With the lifting of the export controls, the access restrictions have now been terminated.<\/p>\n

Anthropic announced that Claude Fable 5 will be made available for global general availability starting Wednesday, July 1, across the Claude Platform, Claude.ai<\/a>, Claude Code, and Claude Cowork. Users on Pro, Max, Team, and select Enterprise plans will be permitted to use Fable 5 for up to 50% of their weekly usage limits at no additional cost until July 7.<\/p>\n

After that date, access will be transitioned to a usage-credit basis. The company is also working to restore access via Amazon Web Services (AWS), Google Cloud, and Microsoft Foundry as quickly as possible.<\/p>\n

Meanwhile, access to Claude Mythos 5\u2014a model designed for defense-oriented cybersecurity work that operates with fewer safety constraints\u2014was restored to select US entities following approval from the US government on June 26.<\/p>\n

The company stated that it continues to work in coordination with the US government to expand access to domestic and international partners under its Glasswing program.<\/p>\n

The cybersecurity report that triggered export restrictions<\/strong><\/p>\n

According to the timeline disclosed by the company, the export controls were instituted on June 12, just three days after the models’ June 9 launch. The regulatory intervention was triggered by a report from Amazon researchers who successfully bypassed the safety guardrails of the Fable 5 model to identify specific software vulnerabilities.<\/p>\n

The report highlighted a single instance where the model generated code demonstrating how the vulnerability in question could be exploited. Anthropic collaborated with the US government and partners, including Amazon, over a two-week period. During subsequent testing, Anthropic confirmed that lower-capacity models\u2014such as Claude Opus 4.8, GPT-5.5, and Kimi K2.7\u2014were also capable of identifying the same vulnerabilities.<\/p>\n

Furthermore, the tests revealed that the code demonstrating how to exploit the vulnerability could be generated by all tested models, including Claude Haiku 4.5, Sonnet 4.6, Opus 4.7, GPT-5.4, and Kimi K2.7.<\/p>\n

Anthropic maintained that the bypass technique did not expose any unique offensive cyber capabilities at the Mythos level, describing it instead as a borderline case involving routine defensive activities.<\/p>\n

Nevertheless, in cooperation with the US government, the company trained an advanced safety classifier targeting the behavior identified in the report, which it claims blocks the exploit with a success rate of over 99%.<\/p>\n

When cybersecurity-related queries directed at Fable 5 are blocked by this system, users will be notified and their prompts will be automatically redirected to the Opus 4.8 model.<\/p>\n

Anthropic acknowledged that the new safety mechanism carries the risk of triggering more frequent false positives\u2014blocking harmless queries during routine coding and debugging\u2014but said it is optimizing the system to reduce these occurrences.<\/p>\n

The company noted that Claude Mythos 5 is more capable of finding and exploiting cybersecurity vulnerabilities than any actor other than high-level human experts, making it an attractive target for malicious actors.<\/p>\n

In contrast, Claude Fable 5 does not possess these offensive cyber capabilities. Anthropic stated that it doubled its staffing resources in the month leading up to the launch to implement its most robust cybersecurity measures to date on the model.<\/p>\n

This approach, termed “defense-in-depth,” combines training the model to reject dangerous requests with the deployment of automated cybersecurity classifiers.<\/p>\n

These classifiers, which function as smaller auxiliary AI systems, detect and block suspicious requests during cybersecurity operations. To counter the risk of “jailbreaking”\u2014where users manipulate the system to bypass safety rules\u2014Anthropic expanded its safety margins significantly.<\/p>\n

While the company acknowledged that this conservative posture leads to the blocking of some harmless queries, it stated that the trade-off was necessary to prevent offensive cyber exploits.<\/p>\n

Researchers at the Center for AI Safety and Innovation (CAISI), an agency under the US Department of Commerce, also evaluated the new measures and confirmed that the cybersecurity guardrails are highly robust.<\/p>\n

A call for industry-wide safety standards<\/strong><\/p>\n

Pointing to a lack of objective, industry-wide standards to evaluate the severity of safety guardrail bypasses, Anthropic announced that it has begun developing a consensus-based framework alongside Amazon, Microsoft, Google, and other Glasswing partners.<\/p>\n

The proposed cybersecurity assessment framework aims to score the severity of vulnerabilities based on four distinct criteria:<\/p>\n