OWASP Top 10 for Agentic Applications 2026
OWASP Gen AI Security Project · v2026 · first published 2025-12-09
Scope
Ten security risks of autonomous and agentic AI applications, ASI01 to ASI10, published by the OWASP Gen AI Security Project (the document is marked "Version 2026" and dated December 2025). A security list: each entry is a class of attack or weakness, with examples, attack scenarios and mitigations. It is not limited to coding agents.
Why it's ingested
OWASP's list of security risks for agentic applications. Having it as a taxonomy node lets readers go from an OWASP entry to the SAICA failure modes it touches. Most entries describe attacks; SAICA modes mostly describe failures an agent makes without an attacker, so only part of the list has a counterpart.
Categories & crosswalks
Each row is one category of this taxonomy with the SAICA failure modes mapped to it and how close each mapping is: “exact” means the same failure, “partial” an overlap, “related” a neighbouring category. 15 mappings cover 9 of the 10 categories; the one with no mapping is listed under them. The category ids and labels were last compared with the source on 2026-11-21.
Mapping drafted by an AI model from the source text; not yet reviewed by a person.
-
ASI01Agent Goal HijackAttackers "manipulate an agent's objectives, task selection, or decision pathways", for example through prompt-based manipulation, deceptive tool outputs, malicious artefacts, forged agent-to-agent messages or poisoned external data. It works because agents "cannot reliably distinguish instructions from related content".
SAICA-KG crosswalks: -
ASI02Tool Misuse and Exploitation"Agents can misuse legitimate tools due to prompt injection, misalignment, or unsafe delegation or ambiguous instruction." The agent stays within its authorised privileges but uses a legitimate tool in an unsafe or unintended way, "for example deleting valuable data, over-invoking costly APIs, or exfiltrating information". Privilege escalation belongs to ASI03 and code execution to ASI05.
SAICA-KG crosswalks: -
ASI03Identity and Privilege AbuseExploits "dynamic trust and delegation in agents to escalate access and bypass controls": delegation chains, role inheritance, control flows and agent context such as cached credentials or conversation history.
SAICA-KG crosswalks: -
ASI04Agentic Supply Chain VulnerabilitiesArise "when agents, tools, and related artefacts they work with are provided by third parties and may be malicious, compromised, or tampered with in transit": models and weights, tools, plug-ins, datasets, other agents, MCP and A2A interfaces, registries and update channels. Examples in the text include typo-squatted tools and endpoints, a compromised MCP or package registry and a compromised npm package installed by coding agents.
SAICA-KG crosswalks: -
ASI05Unexpected Code Execution (RCE)Agentic systems, "including popular vibe coding tools", generate and run code. "Attackers exploit code-generation features or embedded tool access to escalate actions into remote code execution (RCE), local misuse, or exploitation of internal systems." One listed example is "code hallucination generating malicious or exploitable constructs".
SAICA-KG crosswalks: -
ASI06Memory & Context Poisoning"Adversaries corrupt or seed this context with malicious or misleading data, causing future reasoning, planning, or tool use to become biased, unsafe, or aid exfiltration." Context here is what an agent retains, retrieves or reuses (summaries, embeddings, RAG stores), not a one-time prompt.
SAICA-KG crosswalks: -
ASI07Insecure Inter-Agent CommunicationExchanges between agents "lack proper authentication, integrity, or semantic validation", which allows "interception, spoofing, or manipulation of agent messages and intents".
SAICA-KG crosswalks: -
ASI08Cascading Failures"A single fault (hallucination, malicious input, corrupted tool, or poisoned memory) propagates across autonomous agents, compounding into system-wide harm." The entry is about the spread and amplification of a fault across agents, tools and workflows, not about the first fault.
SAICA-KG crosswalks: -
ASI10Rogue Agents"Malicious or compromised AI Agents that deviate from their intended function or authorized scope, acting harmfully, deceptively, or parasitically within multi-agent or human-agent ecosystems." The entry is about the loss of behavioural integrity once the drift has begun, not about the first intrusion.
SAICA-KG crosswalks:
Categories with no SAICA failure mode
This category has no mapping. The reason is the one the crosswalk file states; where it states none, that is said.
-
ASI09Human-Agent Trust ExploitationAdversaries or misaligned designs exploit the trust people place in a fluent, expert-sounding agent "to influence user decisions, extract sensitive information, or steer outcomes for malicious purposes". The risk grows when people approve actions without checking them.
Why it is not mapped: Human-Agent Trust Exploitation: a person is persuaded by an agent to approve or do something harmful. SAICA modes describe what the agent does to code and tasks; none covers manipulation of the human.
Notes
Checked on 2026-11-21 against the PDF "OWASP Top 10 for Agentic Applications 2026" (57 pages, created 2025-12-09) downloaded from genai.owasp.org/download/52117. The ten ids and labels match the document. Descriptions were rewritten from each entry's "Description" section; quoted text is copied from it. Two earlier statements were wrong and are removed: ASI04 does not mention slopsquatting or hallucinated package names (it covers typo-squatted tools, endpoints and packages and compromised registries), and ASI02 is not about exceeding granted authority (that is ASI03). The licence field said "CC-BY"; the document states CC BY-SA 4.0.