The OWASP Top 10 for agentic applications, control by control
OWASP published its first list of risks for AI agents in December 2025. This is what each entry looks like in practice, and the control that answers it.
By the NoFly team
- 01The list
- 02Each risk, and the control for it
- 03Using the list
- 04How NoFly helps
- 05What to do this week
- 06Sources
The list
The OWASP Top 10 for Agentic Applications for 2026 was released on 9 December 2025 by OWASP’s Agentic Security Initiative. It covers systems that plan, call tools, keep memory and hand work to each other, rather than models that only answer questions.
Each risk, and the control for it
ASI01
Agent Goal Hijack
An attacker redirects what the agent is trying to do, usually through content it reads.
ASI02
Tool Misuse and Exploitation
Legitimate tools are used in harmful ways through injection, misalignment or unsafe delegation.
ProtectEvery tool call is decided before it runs against a per-agent allow-list, and risky writes wait for a person.
ASI03
Identity and Privilege Abuse
Delegation chains and inherited credentials are used to reach more than the agent should.
ASI04
Agentic Supply Chain Vulnerabilities
Models, tools, MCP servers or other agent components are malicious or tampered with.
DiscoverInterrogateMCP servers, skills, model files and packages are scanned before they run and matched against advisories, and live servers are questioned for a second face.
ASI05
Unexpected Code Execution
Generated code or tool access turns into code running where it should not.
ProtectShell commands are judged in full before they run, and critical moves such as a script piped from the internet into a shell are refused on the machine.
ASI06
Memory and Context Poisoning
Corrupted memory or retrieved context steers what the agent does later.
ASI07
Insecure Inter-Agent Communication
Messages between agents lack authentication, integrity or validation.
ASI08
Cascading Failures
One fault spreads across agents into harm across the system.
ASI09
Human-Agent Trust Exploitation
People’s trust in the agent is used to get harmful actions approved.
GovernApprovals show the real call rather than the agent’s summary, are bound to that exact request and expire as a no.
ASI10
Rogue Agents
Agents drift from their intended function or scope.
Using the list
Treat the list as a coverage map, not a checklist to tick once. For each entry, ask which control answers it on every surface your agents use, laptops, gateways, hosted platforms and your own code, then attack that control to see whether it holds.
How NoFly helps
Govern
Every rule, finding and control maps to the OWASP Agentic Top 10, alongside the OWASP LLM Top 10, MITRE ATLAS and NIST AI RMF, with controls marked covered, attested, at risk or not assessed.
Attack
Attack scenarios are mapped to the same list, so you can see which entries were tested, which held and which have not been tested yet.
Verify
Controls, including other vendors’, are graded by what they did under attack, so coverage means proven rather than claimed.
What to do this week
- For each of the ten entries, name the control that answers it today.
- Mark the entries where no control exists, or where it has never been tested.
- Pick the highest-risk gap and attack it this month.
More from the blog.
Find out what your AI is allowed to do today.
Book a call and we will map your agents, MCP servers and keys with you, or start on one laptop with the free scan.