What happened: A new IAPS policy memo argues the U.S. needs a strategic response to secure frontier AI systems against theft and sabotage, pointing to cyber-capable models like Anthropic's Mythos and OpenAI's GPT-5.5 as the reason the ‘just ship it' era is over.

Why it matters: If model weights and access programs are strategic assets, they become targets - and ‘security' stops being a slide deck. IAPS calls for standards for high-security data centers, stronger testing with industry, and more government capacity, including empowering NIST's CAISI.

Wider context: The memo lands as labs experiment with restricted-access programs: Anthropic let partners access Mythos Preview to find vulnerabilities, and OpenAI expanded a program giving cyber professionals access to frontier models. The market is drifting toward ‘capability, but with badges and logs.'

Background: IAPS warns risks beyond cyber - including biosecurity - will emerge without additional safeguards. Its recommended posture includes defensive automation, better public-private information sharing, and a third-party evaluation ecosystem. Basically: treat this like infrastructure, not vibes.

IAPS Recommends Federal Action to Secure Frontier AI Models From Theft — PYMNTS

Singularity Soup Take: It's comforting to watch the policy world discover, in real time, that ‘frontier model' is just a fancy way of saying ‘high-value target.' If your competitive edge fits in a server rack, congratulations: you now need boring, expensive, adult supervision.

Key Takeaways:

• Secure the Vaults: IAPS recommends accelerating technical standards for high-security data centers and extending partnership models to include infrastructure housing frontier model weights, so threat detection and sabotage defense are built into the physical layer.
• Defend With Automation: The memo urges scaling defensive automation and building systems to detect and disrupt offensive cyber agents, plus automating defensive R&D across priority safety and security domains - because humans don’t patch at machine speed.
• Build the Risk Hub: IAPS calls for centralised risk information sharing, expanded public-private mechanisms, guidance on agent identifiers for monitoring federal AI use, and investment in evaluation science - with NIST’s CAISI positioned as a key capacity node.