Skip to content
sinc-LLMBook a 15-minute call

OWASP Top 10 for Agentic Applications 2026: ASI01 to ASI10 Explained

The ten agentic AI risks OWASP published for 2026, from Agent Goal Hijack to Rogue Agents: what each covers, how they differ and the controls to start with.

On this page

By Mario Alexandre · November 19, 2026 · 6 min read

The short answer: the OWASP Top 10 for Agentic Applications 2026 is a list of ten security risks specific to AI agents that plan, call tools, keep memory and act across systems. It was published by the OWASP GenAI Security Project's Agentic Security Initiative in December 2025, and its entries run from ASI01 Agent Goal Hijack to ASI10 Rogue Agents. Its central advice is what the authors call least agency: do not give a system autonomy it does not need, because unnecessary autonomy expands the attack surface without adding value. It complements the OWASP Top 10 for LLM Applications rather than replacing it.

The ten risks

ID Risk What it covers, in one line
ASI01 Agent Goal Hijack Manipulated inputs redirect the agent's goals, task selection or multi-step plan
ASI02 Tool Misuse and Exploitation The agent uses a legitimate tool, within its permissions, in an unsafe way
ASI03 Identity and Privilege Abuse Delegation chains, inherited credentials or cached tokens let the agent exceed its intended access
ASI04 Agentic Supply Chain Vulnerabilities Third-party models, tools, datasets, MCP servers, other agents or registries are malicious or tampered with
ASI05 Unexpected Code Execution (RCE) Generated or injected code runs and compromises the host, container or sandbox
ASI06 Memory and Context Poisoning Stored memory, summaries, embeddings or RAG stores are seeded with malicious or misleading data
ASI07 Insecure Inter-Agent Communication Messages between agents are intercepted, spoofed, replayed or manipulated
ASI08 Cascading Failures One fault propagates across agents, tools and workflows into system-wide harm
ASI09 Human-Agent Trust Exploitation People over-trust a fluent agent and approve actions without independent checks
ASI10 Rogue Agents An agent drifts from its intended function and its behavior becomes harmful

Where each risk starts and stops

The document spends real effort on boundaries between entries, which matters when you triage an incident or write a test plan:

  • Goal hijack versus memory poisoning versus rogue agents. ASI01 is an attacker directly altering the agent's goal. ASI06 is the persistent corruption of stored context, which often leads to goal hijack later. ASI10 is misalignment that continues without active attacker control, once the drift has begun.
  • Tool misuse versus privilege abuse versus code execution. ASI02 covers unsafe use of permissions the agent already has, such as deleting valuable data or over-invoking a costly API. If the misuse involves escalation or credential inheritance it belongs to ASI03. If it turns into arbitrary code execution it belongs to ASI05.
  • Cascading failures describe spread, not origin. A tainted dependency, poisoned memory or spoofed message is classified under ASI04, ASI06 or ASI07. ASI08 applies only when that defect spreads across agents, sessions or workflows.
  • Human trust exploitation hides the agent's role. In ASI09 the agent manipulates a person into performing the final, audited action, so the agent's part in the compromise can be invisible to forensics.

Controls to start with for each risk

Each entry has a long mitigation list. These are the first controls from each list that most teams can apply without new infrastructure:

  1. ASI01: treat every natural-language input as untrusted, and require approval through a person, a policy engine or platform guardrails when the agent proposes an action that deviates from the original task.
  2. ASI02: define a least-privilege profile per tool (scopes, maximum rate, allowed destinations) and show a dry run or diff before a delete, transfer or publish is approved.
  3. ASI03: issue short-lived, task-scoped credentials per agent and re-check each privileged step with a central policy engine.
  4. ASI04: pin tools, prompts and configurations by content hash, allowlist dependencies and keep a kill switch that can revoke a compromised component.
  5. ASI05: never run agents as root, ban eval in production agents and separate code generation from execution with a validation step.
  6. ASI06: scan memory writes before they are committed, segment memory by user and domain, and stop the agent from re-ingesting its own outputs into trusted memory.
  7. ASI07: authenticate agents to each other, sign messages and protect exchanges against replay with nonces and timestamps.
  8. ASI08: separate planning from execution through an external policy check, and add circuit breakers and quotas between planner and executor.
  9. ASI09: require explicit confirmation for sensitive actions, keep tamper-proof logs and show a plain-language risk summary rather than a model-generated rationale.
  10. ASI10: keep signed audit logs of agent actions and tool calls, and have a kill switch and credential revocation ready to disable an agent quickly.

The LLM numbers it cites have moved

The agentic list was written against the 2025 edition of the OWASP Top 10 for LLM Applications. The 2026 LLM edition, published in August 2026, renumbered several entries, so check references carefully:

Cited in the agentic list Same entry in the 2026 LLM list
LLM01:2025 Prompt Injection LLM01:2026 Prompt Injection
LLM05:2025 Improper Output Handling LLM10:2026 Improper Output Handling
LLM06:2025 Excessive Agency LLM03:2026 Excessive Agency

The 2026 LLM list states the division of labor from its side: it owns the risk while the model is a component, and the agentic list owns it once the model becomes an actor with tools, memory and downstream consequences. Our summary of the OWASP Top 10 for LLM Applications 2026 covers the other half.

A worked example: an accounts payable agent

The example is fictional. A finance team runs an agent that reads supplier emails, matches invoices to purchase orders, and can draft and send payment requests to the payment system.

What could go wrong Risk Control the team puts in place
An email says "the supplier changed banks, update the account and pay today" ASI01 and ASI09 Bank detail changes are outside the agent's allowed actions; they go to a person with the original supplier record
The agent's API key can also approve payments, not just draft them ASI03 The agent gets a draft-only credential that expires with each run
A new "PDF parsing" tool is added from a public registry ASI04 Only pinned, reviewed tools load; the new tool waits for review
The agent summarizes each supplier thread into long-term memory ASI06 Summaries carry their source email id and are not used as payment evidence
Finance staff approve 40 drafts a day ASI09 Each draft shows the matched PO, the amount difference and the source email, not the agent's explanation

The pattern is the same in every row: the agent may still be fooled, but the action it can take on its own is small, and the evidence a person sees comes from records, not from the agent.

Putting a boundary and a record around your agents

Two scoped projects address the controls that recur across this list. Agent Action Gate deploys a pre-action gate that requires a task to state its starting point, intended result, allowed actions, consequence limit and completion evidence, and checks the gate with synthetic cases in the agreed environment. The published price is $10,000 for 3 to 5 days. Agent Audit Trail sets up a session record that connects assigned instructions to the responsible agent, resulting artifacts, review verdicts and open human decisions. The published price is $12,500 for 4 to 6 days. Final scope and timing are confirmed before work begins. For the approval side, see our guide to human approval workflows for AI agents.

Book a 15-minute call about agent controls · Request an estimate by email

Frequently asked questions about the agentic Top 10

What is ASI01 in the OWASP agentic Top 10?

ASI01 is Agent Goal Hijack: an attacker manipulates the agent's objectives, task selection or decision path through injected prompts, deceptive tool outputs, malicious files, forged agent messages or poisoned external data.

What is "least agency"?

It is the principle the authors add to least privilege and excessive agency: avoid unnecessary autonomy. If a step does not need an agent to decide, a fixed workflow is both cheaper and safer. Our comparison of AI agents and deterministic workflows helps with that choice.

Does the agentic list replace the LLM Top 10?

No. The agentic list builds on the LLM entries and extends them to multi-step plans, tools and delegation. Most agentic systems need both.

Sources

The risks, boundaries and mitigations come from the OWASP Top 10 for Agentic Applications 2026 (version 2026, December 2025, licensed CC BY-SA 4.0) and its publication page; the 2026 LLM numbering comes from the OWASP Top 10 for LLM Applications 2026. The accounts payable agent is a fictional example. OWASP Top 10 for Agentic Applications for 2026 (OWASP GenAI Security Project) · OWASP Top 10 for LLM Applications 2026