Government cyber security agencies in the US, UK and allied countries now say broadly the same thing about AI agents: give them as little access as possible, keep a human approving anything consequential, and assume prompt injection can't be completely prevented. The clearest statement is joint guidance from May 2026 by CISA, the NSA, the UK's NCSC and partners, which advises organisations to use agentic AI only for low-risk, non-sensitive tasks. NIST adds a risk framework and testing research. All of it is written for organisations, but the principles apply directly to anyone running a personal agent.

Who publishes AI agent security guidance?

Body Country Key publication Binding?
NIST and its CAISI US AI Risk Management Framework; agent-hijacking research; 2026 request for information No
CISA and NSA US Joint agentic AI guidance; earlier AI security guidance No
NCSC UK Prompt injection blog; co-author of joint guidance No
ASD's ACSC, Canadian Cyber Centre, NCSC-NZ Australia, Canada, New Zealand Co-authors of joint guidance No

None of this is law. For binding rules, see our guides to US AI agent rules and the EU AI Act.

  1. NIST releases AI Risk Management Framework 1.0
  2. NIST publishes the Generative AI Profile (NIST AI 600-1)
  3. NIST CAISI publishes agent-hijacking evaluation research
  4. UK NCSC: prompt injection is not SQL injection
  5. NIST CAISI requests information on securing AI agents
  6. Joint guidance: Careful Adoption of Agentic AI Services

What does the joint "Careful Adoption" guidance say?

Careful Adoption of Agentic AI Services was published on 1 May 2026 (some sources date it 30 April) by Australia's ACSC, CISA, the NSA, the Canadian Centre for Cyber Security, NCSC-NZ and NCSC-UK. It groups agent risks into five categories: privilege, design and configuration, behaviour, structural, and accountability.

Its core recommendations, in its own words:

  • "strict adherence to the principle of least privilege is critical"
  • "never granting it broad or unrestricted access, especially to sensitive data or critical systems"
  • "organisations should only use agentic AI for low-risk and non-sensitive tasks"

It also recommends mandatory human approval for consequential actions. That matches how consumer agents are built: Meta Muse requires approval for every purchase, Claude Duet asks before sending, paying or deleting by default, and Grok Bot lets you set "Ask first" rules. Our guide to least privilege applies the same idea to personal accounts.

Why does the NCSC say prompt injection may never be solved?

In a December 2025 blog post, Prompt injection is not SQL injection (it may be worse), the UK NCSC explained why the classic fix for injection attacks doesn't carry over. With SQL injection, code and data can be strictly separated. A language model reads instructions and content in the same stream, which makes it an "inherently confusable deputy".

The NCSC's conclusion is sober: "it's very possible that prompt injection attacks may never be totally mitigated", and "The best we can hope for is reducing the likelihood or impact of attacks." It recommends "deterministic (non-LLM) safeguards that constrain the actions of the system", meaning hard limits enforced outside the model.

Vendors say similar things. Anthropic has written that "prompt injection is far from a solved problem", and Meta has acknowledged its defences aren't complete. The hard limits in consumer agents, such as Duet refusing money transfers even with approval, are examples of deterministic safeguards. See indirect prompt injection examples and the lethal trifecta.

What has NIST contributed?

The AI Risk Management Framework

The AI RMF 1.0, released 26 January 2023, organises AI risk work into four functions: Govern, Map, Measure and Manage. NIST says it "is being revised as part of the White House AI Action Plan." The Generative AI Profile (NIST AI 600-1, July 2024) applies it to generative models. Neither is agent-specific.

Agent-hijacking research

NIST's Center for AI Standards and Innovation (CAISI) published research on agent hijacking in January 2025, updated in December 2025, working with the UK AI Security Institute. Testing an upgraded Claude 3.5 Sonnet, a model released in 2024, CAISI found new attacks raised hijacking success from 11% to 81%.

Hijack success with new attacks on a 2024 model (NIST CAISI)
11% to 81%
Average success across five tasks
57%
Success when attacks were retried 25 times
80%

The lesson isn't that a given product is unsafe today, since models have changed since. It is that a defence that looks strong against known attacks can fail against new ones, and that attackers who retry succeed more often.

The 2026 request for information

On 12 January 2026, CAISI issued a request for information on securing AI agent systems (docket NIST-2025-0035, comments due 9 March 2026). It covers indirect prompt injection, poisoned models and specification gaming. Secondary reports describe a related AI Agent Standards Initiative launched in February 2026; we have not confirmed details on a NIST page.

What about CISA's other AI guidance?

CISA has co-authored several AI security publications, including "Deploying AI Systems Securely" (April 2024) and "AI Data Security" (May 2025), and with the UK NCSC the 2023 Guidelines for Secure AI System Development. These predate the agent wave and are aimed at developers and operators. CISA's AI page lists current material; titles may have been updated since we checked.

What should individuals take from this?

The agencies' advice scales down neatly to a personal agent:

  • Least privilege. Connect only the accounts a task needs.
  • Human approval. Keep approvals on for sending, paying, deleting and sharing.
  • Low-risk tasks first. Start with read-only summaries before letting an agent act.
  • Assume injection. Treat anything the agent reads from outside as potentially hostile.
  • Review and revoke. Check the activity log and know how to cut access.

Our safety checklist turns these into specific steps, and the OWASP agentic Top 10 explainer covers the same risks in the security industry's terms. For a primer on the core attack, Get Personal AI Agents explains prompt injection. If you think an agent has been used to defraud you, contact your bank or card issuer, then report it to IC3 in the US or Action Fraud in the UK.