Fort Meade, Md. — The National Security Agency has joined Australia’s cyber authority and several international partners to publish a new cybersecurity information sheet, “Careful Adoption of Agentic AI Services,” aimed at helping operators of critical infrastructure — including in the defense sector — manage the rising risks of autonomous AI systems.
The guidance warns that agentic AI, which can plan and act across interconnected tools without continuous human validation, expands attack surfaces and complicates assurance compared with conventional generative AI. It urges organizations to embed AI security into established cybersecurity practices as the technology matures.
Risk areas highlighted in the document include:
– Excessive privileges that can magnify the impact of a single compromise
– Insecure design and configuration choices that introduce vulnerabilities
– Behavioral issues such as goal misalignment, reward hacking, deceptive actions and unexpected emergent capabilities
– Systemic risks from tightly coupled, interdependent components that broaden the attack surface
– Accountability gaps created by opaque system behavior that hinders auditing and compliance
The report organizes recommended defenses across the lifecycle: designing agents with security in mind, hardening development processes, rigorously vetting third‑party components, deploying with robust safeguards, and operating with strong controls. It advises incremental rollouts, continuous reassessment against evolving threat models, strengthened governance and auditability, persistent monitoring, and maintaining human oversight.
The NSA urged organizations using agentic AI, including those in defense, to review and implement the mitigations.
Co-signatories include the Canadian Centre for Cyber Security, the U.S. Cybersecurity and Infrastructure Security Agency, New Zealand’s National Cyber Security Centre, and the United Kingdom’s National Cyber Security Centre. The report and additional technical guidance are available on the NSA’s website.







