Weekly Digest
Week of 12 Jan 2026
This week at a glance
This week's digest centres on two developments relevant to practitioners working at the intersection of AI capability and governance design. NIST's Center for AI Standards and Innovation has opened a Request for Information on securing agentic AI systems, seeking input on risks specific to autonomous agents—including prompt injection and misaligned objective pursuit—that will inform future voluntary guidelines; given Australian agencies' established reliance on NIST publications, those beginning to govern or procure agentic tools should consider tracking this consultation. Separately, research covered in Import AI raises a longer-horizon policy design question: whether AI regulations could be structured with automated compliance triggers, deferring entry into force until capable enforcement systems exist. Taken together, both items reflect a broader shift in the governance conversation from static rules toward frameworks that account for the dynamic and autonomous character of emerging AI systems.
Headlines
Australian Government1 item
CAISI Issues Request for Information About Securing AI Agent Systems
NIST's Center for AI Standards and Innovation (CAISI) has published a Request for Information seeking input from industry, academia, and security researchers on the secure development and deployment of AI agent systems. The RFI focuses on security risks distinct to agentic AI - including indirect prompt injection, data poisoning, specification gaming, and misaligned autonomous actions - rather than general software vulnerabilities. Responses will inform future voluntary guidelines and CAISI's ongoing research. The comment period closes 9 March 2026. Australian agencies exploring agentic AI use cases or developing related governance frameworks may find the resulting guidance directly applicable.
Key points
- NIST's CAISI has issued an RFI on securing AI agent systems, with submissions closing 9 March 2026.
- The RFI targets risks unique to agentic AI: prompt injection, data poisoning, misaligned objectives, and specification gaming.
- Outputs will inform voluntary US guidelines - a likely reference point for Australian agentic AI governance work.
Implications
- Monitor Agencies with AI strategy or governance functions may want to monitor CAISI's published outputs from this RFI, as resulting voluntary guidelines are likely to be referenced in Australian agentic AI governance discussions.
- Consider Policy and security teams deploying or evaluating AI agent systems could consider whether the RFI's risk taxonomy - prompt injection, data poisoning, misaligned objectives - maps usefully onto existing internal risk assessment frameworks.
Technical Developments1 item
Import AI 440: Red queen AI; AI regulating AI; o-ring automation
Jack Clark's Import AI newsletter (issue 440) covers four research items. Sakana AI's 'Digital Red Queen' paper demonstrates LLM-driven adversarial evolution in a competitive programming environment, with implications for cybersecurity arms-race dynamics. The Institute for Law and AI proposes 'automatability triggers' - regulatory provisions that only activate once AI-assisted compliance tooling exists and meets defined performance thresholds. A University of Toronto/NBER paper applies the o-ring production function to argue that partial automation raises the value of remaining human labour rather than eliminating it. Finally, a multi-institution study finds GPT-4o is equally effective at persuading people toward or away from conspiracy theories, with a system-prompt intervention partially mitigating the bunking effect.
Key points
- Import AI 440 covers four distinct research items: adversarial LLM evolution, AI-automated compliance, o-ring labour economics, and LLM persuasion of conspiracy beliefs.
- The automated compliance piece proposes 'automatability triggers' - regulations that activate only once AI can cheaply enforce them - directly relevant to AI governance design.
- The LLM persuasion research and labour economics item have indirect APS relevance; the adversarial evolution item is primarily technical interest.
Implications
- Monitor AI governance and regulatory policy teams may want to monitor the 'automatability triggers' concept as a potential design pattern for future AI regulation - including any Australian legislative work on mandatory AI standards.
- Consider Agencies developing AI risk frameworks could consider the o-ring labour model when assessing workforce transition impacts of AI deployment within their own operations.
Implications are AI-generated. Starting points, not advice — see methodology for how they're framed.