On February 17, the U.S. National Institute of Standards and Technology announced an AI Agent Standards Initiative, run by its Center for AI Standards and Innovation (CAISI). The goal is standards that let AI agents — which NIST describes as systems capable of autonomous actions — be adopted with confidence, operate securely on behalf of their users, and interoperate across digital systems.
CAISI will work with NIST’s Information Technology Laboratory, the National Science Foundation and other agencies. NIST said research, guidelines and other deliverables will follow in the coming months.
/ 01What NIST announced
The initiative is organised around three pillars:
- supporting industry-led standards for agents and U.S. participation in international standards bodies;
- backing community-led development and maintenance of open-source protocols for agents;
- research on agent security and identity, to make new uses possible and adoption trustworthy.
Two public inputs are already open. CAISI’s Request for Information on AI agent security closes on March 9, and the Information Technology Laboratory is taking comments on a concept paper on AI agent identity and authorization until April 2. NIST said CAISI will also hold listening sessions from April on sector-specific barriers to adopting AI, with a focus on agents.
The security RFI, published in the Federal Register on January 8, is a useful guide to what CAISI is worried about. It defines agent systems as at least one generative AI model plus scaffolding software that gives the model tools for discretionary actions, and notes they may run with little or no human oversight. It says CAISI research “has demonstrated risks of agent hijacking” and names indirect prompt injection — malicious instructions hidden in data the agent reads — as one example of attack. It asks for input on threats specific to agents, including multi-agent systems; on security controls and human oversight; on how to assess agent security; and on how to constrain and monitor the environments agents run in.
/ 02Why it matters
Agents are already being sold into enterprises and government, but the plumbing underneath them is young. How an agent proves who it is acting for, what it is permitted to do, how one agent hands work to another, and how a deployer can tell that an agent was steered by something it read: none of these have settled answers. A federal standards body naming those as priorities gives buyers a vocabulary for procurement and gives vendors a direction for engineering.
NIST does not regulate. Its frameworks are voluntary. But NIST frameworks tend to become a common reference for agencies, auditors and contracts, much as the AI Risk Management Framework has. An agent product that cannot say how it handles identity, authorisation and injected instructions will have a harder time in federal and regulated markets once these documents exist.
/ 03Where knowledge structure fits
Much of the agent security problem comes down to an agent treating everything it reads as equally trustworthy. A web page, an email and an internal policy document all arrive as text in the same context window. Indirect prompt injection works because the model has no built-in way to tell an instruction from a fact, or a trusted source from an untrusted one.
Systems that put a structured layer between raw content and the agent’s actions can help. If documents are first parsed into a knowledge graph — typed entities and relationships, each tagged with its source and provenance — then the agent reasons over facts with known origins rather than over raw text. Policies about what an agent may do can be expressed against the same model: this role may read these classes of record, that action requires this approval. And when an agent does act, the graph gives an auditable record of which facts it used. That does not remove the need for the identity and authorisation standards NIST is pursuing, but it makes them easier to enforce and to check.
Agent security starts with knowing where every fact an agent acts on came from.
/ 04What to watch or do
- Respond while it is open. Organisations deploying agents in regulated settings have until March 9 for the security RFI and April 2 for the identity paper. Concrete failure cases are more useful to standard-setters than general views.
- Protocols. The open-source pillar suggests NIST will engage with the protocols agents already use to reach tools and data. Watch which ones it names.
- Inventory your agents. Before guidance arrives, list what each agent can read, what it can do, and under whose identity. Most of the questions NIST is asking start there.
NIST is right to start with identity and hijacking. An agent that cannot tell a policy document from an instruction buried in an email has no security model worth standardising.
Put a typed graph between raw content and the agent. Each fact carries its source and a trust class; each permitted action is a node tied to a role and an approval; each executed action records the facts it consumed. Indirect injection then has to get past a schema: text from an untrusted web page can populate a low-trust assertion, but it cannot create a permission. When an auditor asks why an agent sent a file, the answer is a query over identity, policy and records. Knowledge Builder enforces semantic constraints over the graph in every pipeline, so an agent that steps outside them fails loudly rather than quietly. Running behind the firewall keeps that record under the deployer’s control.
How we’d model it
Agent action—authorizedBy→PolicyAgent—actsOnBehalfOf→User identityRetrieved fact—hasProvenance→Source documentSource document—hasTrustLevel→Trust class
- Announcing the "AI Agent Standards Initiative" for Interoperable and Secure Innovation — NIST nist.gov ↗
- Request for Information Regarding Security Considerations for Artificial Intelligence Agents — Federal Register (govinfo.gov) govinfo.gov ↗
See how Aktiver turns raw data into knowledge graphs your AI can reason over — deterministic, traceable, behind your firewall.