Urgent.News

What's breaking now, across thousands of outlets.

AI

AI and Machine Learning Security for Government Systems

OMB tells an agency to test a high-impact AI system before it goes live. It does not tell anyone how. That gap is where most of the security risk in federal AI deployments currently sits. OMB Memorandum M-25-21 , issued in April 2025 to replace M-24-10, sets minimum risk management practices for "high-impact" AI: pre-deployment testing, an AI impact assessment, ongoing monitoring, and human…

Federal agencies deploying high-impact AI systems face significant security risks. The Office of Management and Budget (OMB) issued guidelines in April 2025, but lacks detailed instructions on implementation. The National Institute of Standards and Technology (NIST) provides a framework, but does not directly address prompt injection or data leakage in AI systems.

Three main types of government AI systems pose unique threats. Retrieval assistants over agency documents risk indirect prompt injection and data leakage through shared indexes. Predictive models score individuals and can be poisoned through feedback loops. Agents with tool access can invoke unintended actions through LLMs.

The supply chain risk lies in model weights and other components from shared hubs. To mitigate risks, agencies should test AI systems before deployment, focusing on artifacts like model files and permissions. Start by scanning model files for unsafe operators using tools like ModelScan. Require safe formats and pin model revisions to prevent unauthorized changes.

Test the AI system's permission boundaries with a canary string, planting it in restricted documents and checking if unauthorized users can access it. Monitor retrieval traces and logs to track what the model sees, ensuring it respects access controls. Implement logging that captures user information, model details, retrieved document IDs, tool calls, and outputs. This creates a comprehensive audit trail in line with SP 800-53 Rev. 5 controls.

However, pre-deployment testing alone is insufficient. Results may not be applicable to future model versions or changes in the AI system's supply chain. Maintain version notices and run tests on each update. The canary test confirms the presence of security vulnerabilities, not their absence. Continuous monitoring and testing are crucial to ensure AI systems align with the minimum risk management practices outlined by OMB and NIST.

Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.

Read the original at dev.to →

More in AI

Two AI poker bots announced a full house. Only one had it.

I run aipoker.bot, a small site where you sit down and play poker against a lineup of AI poker bots. Some of them are language models, some are old-school rule engines like Slumbot, Shanky, Warbot and…

  • One AI bot announced full house with tens, another correctly identified its full house
  • Bots' hand descriptions often mismatched with actual cards held
  • Discrepancy highlights gap between narrative and reality in AI poker bots

More from Monday 16 November →