← Frameworks / OWASP Agentic Top 10 2026 / Coverage Analysis

OWASP Top 10 for Agentic Applications 2026 โ€” SP 800-53 Coverage

How well do NIST SP 800-53 Rev 5 controls address each OWASP Agentic Top 10 2026 requirement? This analysis maps from framework clauses back to SP 800-53, with expert coverage weightings and gap identification.

Clauses: 10
Avg Coverage: 59.5%
Publisher: OWASP GenAI Security Project
Coverage Distribution
Full (85-100%): 0 Substantial (65-84%): 5 Partial (40-64%): 5 Weak (1-39%): 0

Clause-by-Clause Analysis

Sorted by clause
ASI01 Agent Goal Hijack

Rationale

SI-10 validates prompts, documents and retrieved content before they can steer the agent's plan. AC-04 filters what flows in from email, files, web pages and other agents. AC-06 keeps the agent's tools to least privilege, so a hijacked goal can do less harm. AC-03 enforces approved actions, and AC-03(02) puts a person's approval in front of high-impact or goal-changing ones. CM-03 brings system prompts and goal definitions under change control. SI-04 compares goal state and tool use with a baseline and alerts on drift, and AU-02 logs each deviation. CA-08 red-team exercises simulate a goal override. PM-12 adds agents to the insider threat programme.

Gaps

The root cause is a model that can't reliably tell instructions from content. The risk also needs each planned action checked at run time against the user's original task, with the declared goal and its limits bound to every step. Spotting drift needs the active goal kept as tracked state. SI-10 checks inputs against a defined syntax, so detecting instructions hidden in natural language calls for measures of your own.

ASI02 Tool Misuse and Exploitation

Rationale

AC-06 and CM-07 give each tool least privilege and only the functions its task needs, and CM-07(05) keeps an allowlist of approved, version-pinned tools. AC-24 decides on each tool call before it runs, and AC-03 enforces the decision, holding destructive actions for a person's confirmation through AC-03(02). SI-10 validates arguments against the tool's schema, so raw model output never reaches a shell or a database. IA-13 issues short-lived tokens scoped to the tool and the session. SC-39 sandboxes tool execution, and SC-07 lets outbound traffic reach approved destinations only. SC-06 quotas cap rate and token use. SI-07 checks pinned tool descriptions for tampering. SI-04 flags unusual tool chains, such as a database read followed by an external send. AU-02 logs every call, and AU-09 protects the log from change.

Gaps

The risk also needs a check that each permitted call serves the user's task, judged by its meaning as well as its syntax. Policy needs to see sequences, since tools approved one by one can still form a harmful chain. Tool names need to resolve safely, failing closed on look-alike or ambiguous names. A tool description poisoned at its source passes an integrity check. People approving an action need a dry run or a preview of its effect. Spend limits fit SC-06 only loosely, since SC-06 protects availability.

ASI03 Identity and Privilege Abuse

Rationale

AC-02 manages each agent as its own account, and AC-02(12) flags atypical use such as newly gained scopes. AC-06 keeps permissions to the task, so a manager agent can't pass its full access to a narrow worker. AC-24 decides on each privileged step, so an agent re-checks relayed instructions before acting on them. AC-03(08) revokes an authorisation when its attributes change, and AC-03(02) puts a person's approval in front of high-privilege actions. IA-02 ties each action to the user the agent acts for. IA-09 authenticates agents as services, which defeats forged agent identities. IA-13 issues short-lived, audience-restricted tokens and verifies them. IA-05 manages agent credentials, and IA-05(13) bars reuse of cached ones. IA-11 asks for re-authentication on a context switch. SI-14 starts each session from a known state and wipes it at the end, and SC-04 stops data held for one session reaching the next. SI-04 detects requests for new scopes, token reuse outside the original task and device-code phishing.

Gaps

The risk also needs delegation handled as a chain: scope narrows at each hop between agents, and the original user's intent is re-checked before one agent acts on another's behalf. Tokens need binding to a signed intent that names subject, audience, purpose and session, which goes beyond audience restriction. SP 800-53 knows users, devices and services, so an agent's persona and its entry in an agent registry call for measures of your own.

ASI04 Agentic Supply Chain Vulnerabilities

Rationale

CM-08 keeps an inventory of AI components, the job of an AI bill of materials, and SR-04 records the provenance of models, tools, prompts and datasets. SR-05(02) assesses each component before it is accepted or updated. CM-07(05) allows approved, pinned components only, and CM-14 blocks anything without an approved signature. SR-11 detects counterfeits, such as a typosquatted package or a server posing as a known tool. SI-07 checks signatures and hashes at install and again at run time, and acts on drift. CM-03 puts prompts, orchestration scripts and memory schemas under version control with review. IA-09 authenticates other agents and tool servers before connecting to them. SA-09 holds providers of outside agents, tool servers and registries to the organisation's requirements. SC-39 runs sensitive agents in sandboxes, and SI-04 watches components for anomalies at run time. IR-04 supplies the kill switch that disables a compromised tool, prompt or agent connection wherever it runs.

Gaps

SP 800-53's supply chain controls assume components are chosen and installed ahead of time. Agents also load tools, prompt templates and peer agents while a task runs, and that live assembly needs checks of its own. Tool descriptions and agent cards need inspecting for hidden instructions. An AI bill of materials needs to record models, prompts and datasets as well as code. Registries of agents and tool servers need a basis for trust, such as signed agent cards.

ASI05 Unexpected Code Execution (RCE)

Rationale

SI-10 validates and encodes generated code and its inputs, and SI-10(06) blocks injection such as a shell command hidden in a file name. CM-07 bans unsafe functions such as eval in production agents, keeps an allowlist for automatic execution and confines unvetted code. SC-18 treats code fetched or generated at run time as mobile code and stops it running automatically. SC-39 gives each session its own sandbox, and SC-07 limits the sandbox's network access, which blocks reverse shells. AC-06(08) keeps code from running with more privilege than the user who invoked it. CM-05(05) keeps agents from changing production directly. AC-03(02) puts a person's approval in front of elevated runs. SA-11 scans and tests generated code before it runs. CM-11 governs the packages an agent installs, and SI-03 scans them for malicious code. SI-04 watches execution at run time, and AU-02 logs each generation and run.

Gaps

Code written at run time skips the development pipeline that SA-11 assumes. The risk needs model output tracked as tainted all the way to execution, with a validation gate between writing code and running it. It also needs a way to spot hallucinated code that carries a backdoor. Agent memory systems need safe evaluators in place of eval.

ASI06 Memory & Context Poisoning

Rationale

SI-10 scans every write to memory, the agent's own output included, before it is committed, and SI-10(05) admits authenticated, curated sources only. AC-03 and AC-06 give each task least-privilege access to memory stores. AC-04 keeps flows apart by session and domain, and stops the agent's own output flowing straight back into trusted memory. SC-04 keeps one tenant's memory from reaching another. AC-16 attaches provenance, trust scores and tenant tags to entries, so retrieval can weigh or exclude low-trust content. AU-10(02) binds each entry to whoever produced it. SC-28 protects memory stores at rest, and SI-07 detects unauthorised changes to them. SI-04 flags suspicious updates and unusual write rates. SI-21 expires information that is unverified or no longer needed, which limits how long poison lasts. CP-09 snapshots let a poisoned store roll back. CA-08 tests the paths by which memory can be poisoned.

Gaps

SP 800-53 treats an agent's memory as ordinary stored data. The risk also needs ways to catch poison that is well formed and arrives through normal channels, down to the level of embeddings. Retrieval needs to weigh entries by trust and let low-trust ones decay, with two signals required before high-impact memory surfaces, such as a provenance score and a person's check. Slow drift in what an agent remembers needs watching across sessions. Re-ingesting the agent's own output needs a rule of its own, beyond AC-04's general flow rules.

ASI07 Insecure Inter-Agent Communication

Rationale

IA-09 authenticates each agent with its own credentials before any exchange, discovery and coordination messages included. SC-17 issues agent certificates under a policy and trusts approved anchors only, the basis for pinning and signed agent cards. SC-08 encrypts and integrity-protects messages end to end, and SC-08(04) hides traffic patterns. SC-23 protects sessions against hijacking and replay, with identifiers tied to the task. SC-16 binds sender, audience and task context to each message and protects them from forgery. AU-10 binds each message to the agent that sent it. AC-04 sets which agents may exchange which messages. SI-10 validates messages against typed, versioned schemas. CM-07 turns off weak and legacy modes, and SC-07 gateways enforce protocol versions and formats, which together stop downgrade attacks. SC-31 analyses covert channels through which agents could leak or infer data. SI-04 watches for unusual routing and discovery traffic.

Gaps

The risk also needs checks on what each message means. An authenticated, well-formed message can still carry hidden instructions or a tampered goal, so its intent needs comparing with the one first sent. Agents can also read one instruction as different intents. Discovery needs attested registries and signed agent cards. The protocols agents speak, such as the Model Context Protocol (MCP) and Agent2Agent (A2A), need rules of their own.

ASI08 Cascading Failures

Rationale

SA-24 designs the system to stay resilient when a model, an agent or an outside service fails. SC-07 segments the network and isolates agent components, so a fault stays inside its trust boundary. AC-06 and scoped APIs limit how far a faulty agent can reach. SI-15 validates each agent's output before it passes downstream, and AC-04(09) holds high-risk outputs for human review. SI-04 detects rapid fan-out, feedback loops and drift from decision baselines. CP-12 drops the system into a restricted safe mode when trouble starts, which works as a circuit breaker, and SC-06 quotas cap the blast radius. CM-04 tests a policy change in a separate environment before it goes live. AU-09 keeps logs of messages and decisions tamper-evident, and AU-10 ties each logged action to the agent that took it, so a cascade can be traced.

Gaps

The risk needs propagation treated as a failure in its own right, such as fan-out across agents and loops between them. Every action that spreads needs lineage data showing where it came from. Faults can spread faster than people can review them, so circuit breakers between planner and executor need to act on their own. Human oversight tends to weaken after a run of successes, which calls for watching approval habits over time. CM-04 covers impact analysis in general, and replaying recorded agent actions in a copy of production before a policy change goes further.

ASI09 Human-Agent Trust Exploitation

Rationale

AC-03(02) requires a second, independent approval before risky actions, such as a payment to new bank details. AC-04 blocks network and state-changing calls from a preview, so a read-only preview has no side effects. AC-16 attaches source and time to recommendations and outside data, and SI-07 checks their hashes and signatures, so content without trusted provenance can be blocked. AU-10 binds each recommendation to the agent that made it, and AU-09 protects the records, so the agent's part in an action a person approved stays visible. SI-04 watches for sensitive data in conversations and for actions that stray from approved workflows. IR-06 gives users a way to report manipulative behaviour, and IR-04 can lock down an agent's capabilities while a report is reviewed. AT-02 teaches users about manipulation and the limits of agents, and AT-03 trains the people who oversee them.

Gaps

SP 800-53 treats people as trained operators of a system. In this risk they are the target, and most of what the risk needs lies in human behaviour. It needs defences against automation bias and against emotional or human-like persuasion by the agent. Explanations need to reflect the agent's real reasoning. Interfaces need confidence cues and risk badges that prompt a second look at high-impact actions, with risk summaries written by the platform instead of the model. The agent's autonomy, and the oversight it gets, need to follow the risk of each situation.

ASI10 Rogue Agents

Rationale

AC-06 keeps API scopes narrow, with no access to provisioning APIs, so a drifting agent can't copy itself or delete backups. SC-39 runs agents in restricted environments, and SC-07 sets trust zones with strict rules between them, where SC-07(20) can isolate a suspect agent. IA-09 gives each agent a cryptographic identity, so other agents can reject an impostor posing as a monitor or an approver. SC-12 keeps signing keys in a hardware module or key service, with rotation and revocation, so agents never hold long-lived keys. SI-04 runs behavioural detection for collusion, coordinated false signals and excess actions. AU-02 logs every action, tool call and message between agents, AU-09 keeps the log immutable, and AU-10 binds each entry to the agent that acted. IR-04 provides kill switches and quarantine for forensic review, and CP-10 restores a remediated agent to a trusted baseline before it rejoins production.

Gaps

SP 800-53 assumes a process behaves as designed until someone compromises it. A rogue agent can drift while it holds authorised access. The risk needs checks on behaviour as such, covering goal drift, scheming, reward hacking and collusion between agents. Each action needs checking against a signed manifest of the agent's declared tools and goals, and the agent's behaviour needs testing at intervals with challenge tasks. Self-replication needs a control of its own.

Methodology and Disclaimer

This coverage analysis maps from OWASP Agentic Top 10 2026 clauses/requirements back to NIST SP 800-53 Rev 5 controls, assessing how well the SP 800-53 control set addresses each framework requirement.

Coverage weighting represents an informed estimate based on control-objective alignment, not a definitive compliance determination. Weightings consider whether SP 800-53 controls address the intent of each framework requirement, even where terminology and structure differ.

This analysis should be validated by qualified assessors for use in compliance or audit activities. The authoritative source for any compliance determination is always the framework itself.