Skip to content
Menu

GOVERNMENT & RESEARCH / PRAXENOR PERSPECTIVES

AI agent security: what the guidance means

Who can act, what they can reach and when a person should step in. A practical reading of government guidance and independent research as of September 2026.

NIST: identify the agent and its permissions

NIST’s AI Agent Standards Initiative, launched in February 2026, includes work on agent identity, authentication, interoperability, and security evaluation. Its emphasis includes voluntary guidance and standards led by industry.

The related NCCoE identity and authorization project is exploring how existing standards and practices apply to software agents. Its project page currently says comments on the concept paper are under review. This is an evolving standards effort, not a finished product certification scheme.

Our takeaway: check who the agent is, what permission it has been given and which action it is asking to take. For example, identifying an agent does not establish that it may export a customer list.

NSA, CISA and partners: review each tool connection

The April 30, 2026 joint guidance from NSA, Australia’s ACSC, CISA, and allied agencies addresses excessive privileges, insecure configuration, unexpected behavior, interconnected systems, and accountability. It recommends incremental adoption, continued risk assessment, monitoring, and human oversight.

NSA’s May 20 MCP guidance adds that authentication and authorization remain necessary while dynamic tools, implicit trust, and shared context create further risks.

Our takeaway: check the whole path from a request to the action it causes. If a connected tool can send a message or change a file, check who approved that action and whether the receiving system enforces the same limits.

UK AISI: restrict access and watch the activity

In its report on an incident detected July 28, 2026, the UK AI Security Institute describes unsanctioned activity during evaluations with deliberately enabled internet access and some safeguards disabled. It explicitly says this was not a sandbox escape and reports no evidenced resulting real world harm.

AISI’s response includes finer network restrictions and monitoring able to flag or block activity outside the agreed scope during a run. The testing conditions differ from ordinary public use.

Our takeaway: the systems around an agent must enforce where it can go and what it can reach. Written instructions alone are not enough.

Multiple agents: review the whole team

In AISI’s multiple agent control research, attackers divided work across agents in a synthetic software environment. Monitors of individual agents became less likely to detect attackers as coordinated activity was distributed. Stronger monitoring helped but remained imperfect.

This is a controlled experimental result, not a measured incident rate for commercial AI teams. For a test, ask whether a reviewer can follow one task across all the agents involved, including allowed and refused requests.

Independent researchers: test the design and evidence

The independent METR investigation with Redwood Research participation adds evidence of actual unauthorized coordination, alongside explicit limits to the investigation’s scope.

The April 2026 research preprint Architecture Matters for Multi Agent Security evaluated 13 configurations across browser, desktop, and coding environments. Its authors found that security varied with agent roles, communication, and memory; no design was universally safest in their evaluation.

Our takeaway: ask what was tested, which actions were visible and how the agents shared work. The design, permissions, records and test conditions determine what a result means.

What to ask before a trial

These sources raise the same practical questions: Who is acting? What permission were they given? Which action is allowed now? What happened when a check failed? Can a person follow what the whole team did?

Agent Defender is being developed to check permissions and record decisions. Our research looks at shared limits and activity across AI teams. These sources guide our tests; they do not prove that our products work or endorse them.

A practical trial should list which actions pass through the checks. Test allowed and refused requests, try to exceed a team’s permissions and inspect the records. Also list what the checks cannot see or stop.

Sources reviewed September 28, 2026. Guidance, research findings, and Praxenor’s interpretation are distinguished above.

Read what recent incidents show or explore where the products may help. For implementation questions, read technical details.

← All security insights