IAPS Urges Federal Action to Secure Frontier AI Models

PYMNTS reports that the Institute for Applied Policy Studies (IAPS) issued a May 14 memo recommending federal action to protect frontier AI models from theft, sabotage, and other security threats. The memo, summarized by PYMNTS, cites the cyber capabilities of Anthropic's Mythos and OpenAI's GPT-5.5 as context and includes the line, "Risks beyond cyber, including biosecurity threats, are likely to emerge without additional safeguards," attributed to IAPS researchers in the memo.

PYMNTS reports the IAPS memo recommends several concrete steps for federal agencies, including:

• •Accelerating development of technical standards for high-security data centers, per the memo reported by PYMNTS.
• •Coordinating with industry to test advanced security measures for model infrastructure, per PYMNTS.
• •Extending partnership models to include infrastructure that houses frontier model weights for voluntary threat detection and analysis, per PYMNTS.
• •Expanding intelligence collection on adversarial development and use of frontier AI, according to PYMNTS.
• •Scaling defensive automation and building infrastructure to detect and disrupt offensive cyber agents, as stated in the memo summarized by PYMNTS.

PYMNTS also reports operational context: Anthropic gave select partners early access to Mythos previews so partners could test defenses, and PYMNTS reports Anthropic said its Mythos Preview had "ferreted out" more than 10,000, per a May 22 statement. PYMNTS further reports that OpenAI said on April 14 it was expanding a program that gives outside testers access to frontier models.

Industry observers often note that protecting model weights and access points requires a blend of physical security, enclave-style compute, cryptographic controls, and continuous monitoring. Organizations building and operating large models typically layer threat detection, red-teaming, and controlled partner programs to surface vulnerabilities before adversaries can exploit them. Scaling defensive automation and automating defensive R&D, as the IAPS memo recommends, aligns with broader industry moves to reduce mean time to detection for model-targeted attacks.

Observers tracking national-security and AI policy debates see proposals like those in the IAPS memo as part of a pattern where capability advances from major model providers prompt calls for government-industry coordination on custody, access, and supply-chain protections. Reporting by PYMNTS places these recommendations alongside contemporaneous provider efforts to invite outside testing of frontier systems.

For practitioners and procurement teams: look for federal requests for comment, draft technical standards for high-security model hosting, public-private testing initiatives, and expanded intelligence sharing mechanisms. For security engineering teams: expect increased emphasis in guidance on protecting model weights, logging and provenance of model calls, and automated defensive playbooks.