Anthropic has constructed an artificial intelligence system so capable that executives refuse to release it to general users. The company revealed on Wednesday that its Claude Mythos Preview demonstrates unprecedented autonomous capabilities in cybersecurity analysis.
This announcement arrives merely weeks after the Trump administration barred federal agencies from utilizing Anthropic products for six months. Washington officials had accused the firm of attempting to influence Pentagon procurement decisions while compromising national security interests.
Department of Defense officials subsequently finalized an agreement with competitor OpenAI to implement their technology within classified military networks.
Government Tensions And Strategic Shifts
The Trump administration restriction emerged following allegations that Anthropic had improperly pressured defense officials regarding contract awards. Officials claimed these actions endangered national security while prompting the Pentagon to seek alternative suppliers for artificial intelligence tools.
Current discussions between Anthropic executives and federal agencies represent a notable shift in the company’s relationship with Washington regulators. These conversations focus on evaluating whether the advanced system merits special classification or restricted access protocols beyond standard commercial restrictions.
Autonomous Capabilities Prompt Industry Alarm
Logan Graham leads Anthropic’s frontier red team and described the system as follows.
extremely autonomous
The system can identify tens of thousands of software vulnerabilities without human assistance. Graham emphasized that unlike earlier generations, this model generates functional exploits rather than merely detecting weaknesses.
Previous Security Incidents
Internal documentation regarding the unreleased model appeared in public data repositories during February after employees misconfigured access permissions. Thousands of sensitive documents remained exposed to internet users for an unspecified duration before security teams discovered the configuration mistake.
Earlier this month, engineering staff accidentally published over five hundred thousand lines of proprietary code containing unreleased features.
Corporate spokespersons characterized this incident as follows.
human error, not a security breach.
These repeated lapses have intensified scrutiny regarding the company’s ability to safeguard technologies far more dangerous than leaked documents.
Controlled Deployment Through Project Glasswing
Anthropic refuses to distribute the model through standard commercial channels until comprehensive safeguards limit its capacity for autonomous exploitation. Instead, the company established Project Glasswing, an initiative involving more than forty major technology corporations and government contractors.
Participating organizations include Apple, Amazon, Microsoft, Google, and NVIDIA, representing unprecedented collaboration among industry giants. These partners will test the system’s ability to identify and remediate vulnerabilities within critical infrastructure software without causing accidental harm.
Access remains restricted to select cybersecurity firms and technology companies meeting stringent operational security requirements.
Risk Assessment And Future Implications
External evaluation teams including federal agencies have examined potential hazards across multiple categories including cybersecurity, loss of control, and CBRN threats. Anthropic incorporated these findings into comprehensive risk assessments before determining that public release would pose unacceptable dangers to global security.
Graham told the New York Times that this moment represents the following perspective.
the starting point for what we think will be an industry change point, or reckoning, with what needs to happen now.
The company insists that artificial intelligence with such destructive potential requires new governance frameworks before any broader deployment can proceed safely. Washington officials continue evaluating whether the technology warrants classification as a controlled defense article or dual-use technology subject to export restrictions.