HYPERNOVELTYINSTITUTE

Technical intelligence
for accelerating change.

Search the record

Home/Analysis

Daily Hypernovelty Lead · November 11, 2026

The Agent’s Attack Surface Is Everything It Trusts

On November 9, 2026, the National Institute of Standards and Technology released an initial public draft of NIST SP 2400-102e2026, a taxonomy focused on attacks against agentic software systems. Comments are open through December 18. The document is voluntary guidance. It sets no legal requirements and does not prescribe an organization’s risk tolerance. Its vocabulary points toward a practical shift for anyone deploying AI agents: the security surface includes the full sequence of what an agent does, even when its final answer looks fine.

Navy operations room
Photo: Petty Officer 1st Class Samuel Souvannason / U.S. Navy · Public domain · Cropped

What NIST means by an agentic system

The draft defines agents as systems that interact with an environment, receive information, and execute actions. That interaction layer includes tools such as browsers, code interpreters, file systems, and APIs. It includes memory that can persist across sessions. It includes the ability to communicate with other agents. The document divides the workflow into four phases: task instructions, planning, execution, and observation. That four-part loop is the frame NIST uses to map where attacks can enter, and attacks can enter at every stage. As the draft states directly, "these attacks can be mounted across the entire agentic lifecycle, spanning task instructions, planning, execution, and observation."

This framing shifts the question from "did the agent produce a correct answer?" to "what did the agent do, and what did it trust, while getting there?"

Resources the agent trusts become the attack surface

The draft gives particular attention to resources an agent treats as authoritative. Tools can execute with the agent’s privileges and observe its working state. Skills, which are reusable task-guidance modules, may be followed as trusted instruction. Peer agents can enter a workflow as trusted parties. NIST describes malicious tools, skills, and agents as an especially strong form of control over the resources an agent uses.

The trust boundary extends beyond the model to everything the agent accepts as reliable input. A tampered tool, a skill from an unvetted source, or a peer agent operating under compromised instructions can redirect behavior without leaving an obvious mark on the final output. The draft’s supply-chain section describes tool poisoning through malicious instructions in metadata. It also describes a “rug pull” that changes tool descriptions after installation, along with skills that may persist across sessions as trusted guidance.

Verification bottleneck

The draft notes that trajectory-level monitoring, which observes the full sequence of an agent’s interactions rather than individual requests, may detect decomposition attacks whose individual steps each appear benign. A decomposition attack breaks a harmful objective into small, unremarkable actions that pass inspection individually but accumulate into something damaging. Detecting that pattern requires a view of the path, rather than the destination alone.

The document also notes that agents can accumulate memory, permissions, tool access, and operational trust across sessions. Checking one exchange in isolation may miss failures that only become visible across a longer arc of activity. Persistence changes the security problem because yesterday’s compromised memory can influence tomorrow’s work.

NIST is candid about the state of the field: “These mitigations remain at an early stage.” Policy guardrails must anticipate harmful actions, while adaptive attacks have bypassed many published defenses against jailbreaks and prompt injection. Trajectory monitoring is a promising direction, but the draft does not present it as a finished safety net. Deployment is moving faster than the controls designed to supervise it.

What a team can reconstruct after the fact

A practical test is to ask what a team can reconstruct after an agent has acted: the instructions it received, sources it retrieved, tools it called, permissions it exercised, memory it read or wrote, other agents it contacted, actions it attempted, and moments when a person could have paused or reversed a step.

If most of that trace is unavailable, the team is effectively performing output-only review. That can catch visible failures. It can miss cases where the route to a plausible answer was itself the problem. Agent hijacking, memory poisoning, and malicious skill loading can all alter the working process before a reviewer sees the result.

From Hypernovelty’s reading, the unit of trust is moving from the model’s response toward the agent’s working environment and action history. NIST does not state that conclusion in those terms. It follows from the structure of the taxonomy: influence can enter through instructions, planning resources, tools, memory, execution feedback, or other agents. The visible answer is one artifact from a much larger process.

This concern would weaken if output checks reliably caught attacks assembled across many individually harmless steps. Mature trajectory controls that hold up when attackers adapt to them would also change the picture. Until then, a clean answer is evidence about the destination. It says much less about the route.


Sources

Back to the archive