Taxonomy of Failure Mode in Agentic AI Systems (Microsoft AI Red Team, version 1.0)
Microsoft AI Red Team (AIRT) · v1.0 (2025-04) · first published 2025-04-24
Scope
Version 1.0 of the Microsoft AI Red Team's list of failure modes of agentic AI systems: 27 modes on two axes, safety or security and novel (only in agentic systems) or existing (known from other AI systems, with higher risk in agents). 4 novel safety, 6 novel security, 7 existing safety and 10 existing security modes. The paper calls its method a failure mode and effects analysis of current and envisaged systems. It is about agentic systems in general, with many multi-agent cases, not about coding agents.
Why it's ingested
A vendor red team's list of how agentic systems fail, widely cited in security discussions of agents. Several of its modes are close to SAICA failure modes (Excessive agency, Hallucinations, Function compromise and malicious functions, Memory poisoning); most of the rest concern multi-agent attacks and responsible-AI harms that SAICA does not cover. See data/crosswalks/msft-airt-to-saica.yml.
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. 13 mappings cover 11 of the 27 categories; the 16 with no mapping are 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.
-
MS-NF-04Prioritization leading to user safety issues"The autonomy granted to agentic AI systems may lead to the system prioritizing its given objective ahead of the safety of users or other systems unless systems are given robust safety guardrails." Sample scenario: an agent that must add database entries deletes all existing entries to make room.
SAICA-KG crosswalks: -
MS-NS-05Agent provisioning poisoning"The manipulation of the method by which new agents are deployed to introduce a malicious element to new agents deployed to the system, or to deploy a specifically malicious agent."
SAICA-KG crosswalks: -
MS-EF-06Hallucinations"The agent produces responses that include incorrect information presented factually by the agent."
SAICA-KG crosswalks: -
MS-EF-07Misinterpretation of instructions"The agent-wide latitude of actions and roles leads it to misinterpret the user’s intent and perform incorrect actions." Sample scenario: asked to "get rid of it" about one record, the agent deletes the whole table.
SAICA-KG crosswalks: -
MS-ES-01Memory poisoning and theft"A threat actor can manipulate future actions of an agent by adding content, most notably malicious instructions, to the system’s memory, which the agent will process each time it is recalled." The details section is headed "Memory poisoning" and does not describe theft.
SAICA-KG crosswalks: -
MS-ES-02Targeted knowledge base poisoningWhen agents have access to knowledge sources specific to their role, often through RAG, "there is an opportunity for a threat actor to poison these knowledge bases with malicious data".
SAICA-KG crosswalks: -
MS-ES-03Cross domain prompt injection (XPIA)"Due to an agent’s inability to distinguish between instructions and data, any data source that the agent ingests that includes instructions poses a risk of that instruction being actioned by the agent, regardless of its provenance." The overview table lists this mode as "XPIA"; the paper notes that others call it indirect prompt injection.
SAICA-KG crosswalks: -
MS-ES-05Function compromise and malicious functions"A threat actor compromises a tool or function available to an agent to use the invocation of that plugin or response from it to manipulate the agent or perform a malicious action." The details section is headed "Tool compromise".
SAICA-KG crosswalks: -
MS-ES-06Incorrect permissions"The agent is provided with permission to access data or perform actions that are above and beyond those provided to an end user." Workflow logic issues let a threat actor use this to reach data or actions they should not have.
SAICA-KG crosswalks: -
MS-ES-08Insufficient isolation"An agent that can perform an unstructured action does so in a way that allows it to interact with systems, users, or components outside of the intended scope of the agent."
SAICA-KG crosswalks: -
MS-ES-09Excessive agency"An agent is provided with insufficient scoping and direction in its actions, resulting in it making decisions and taking actions beyond what is expected of the agent." Sample scenario: asked for advice on an underperforming employee, an agent with full access to the HR system terminates the employee.
SAICA-KG crosswalks:
Categories with no SAICA failure mode
These 16 categories have no mapping. The reason is the one the crosswalk file states; where it states none, that is said.
-
MS-NF-01Intra-agent Responsible AI (RAI) issues"Communications between agents in a multi-agent system could include RAI harms that are either exposed to the user during output or are included in transparency logging."
Why it is not mapped: Intra-agent Responsible AI (RAI) issues: harmful content in messages between agents. No SAICA mode covers content harms.
-
MS-NF-02Harms of allocation in multi-user scenariosAgents with autonomy have to balance priorities, for example when scheduling meetings. "If not addressed in system design, biases in the LLMs used by agents could lead to different users or groups being prioritized differently, leading to differing quality of service."
Why it is not mapped: Harms of allocation in multi-user scenarios: unequal service between users. No SAICA mode covers fairness.
-
MS-NF-03Organizational knowledge loss"A subset of overreliance." An organization that delegates significant powers to agents "could see a breakdown in knowledge or relationships as interactions are handled agent-agent".
Why it is not mapped: Organizational knowledge loss: an effect on the organisation, not a failure of one agent run.
-
MS-NS-01Agent compromise"The compromise of an existing agent with new threat actor–controlled instructions, or a threat actor–controlled model that breaks existing guardrails and introduces a malicious element to an agent or the system."
Why it is not mapped: Agent compromise: an attacker takes control of an agent. SAICA modes describe failures the agent makes unforced; there is no takeover mode. The nearest rows are the prompt-injection ones at MS-ES-03.
-
MS-NS-02Agent injection"The introduction of new malicious agents into an existing multi-agent system with the intent of performing a malicious action or having a detrimental impact on the system." This is the injection of agents, not prompt injection (see MS-ES-03).
Why it is not mapped: Agent injection: malicious agents are added to a multi-agent system. No single-agent counterpart. (Earlier rows read this as prompt injection, which is MS-ES-03.)
-
MS-NS-03Agent impersonation"The introduction of a new malicious agent into a system that impersonates an existing agent in a way that is accepted by other agents in the system."
Why it is not mapped: Agent impersonation: a malicious agent takes the name of an existing one. Similar in spirit to typosquatting, but nothing is installed or invoked from a package source.
-
MS-NS-04Agent flow manipulation"A threat actor compromises some part of the agentic AI system to subvert the flow of the agent system." The flow can be ended, redirected or altered, through crafted prompts, a compromised framework or the network.
Why it is not mapped: Agent flow manipulation: an attacker redirects or ends the flow of a multi-agent system. No SAICA mode.
-
MS-NS-06Multi-agent jailbreaksIn a multi-agent system "a jailbreak could be generated as part of multiple agent interactions, resulting in Agent Compromise while avoiding jailbreak detections".
Why it is not mapped: Multi-agent jailbreaks: a jailbreak assembled across several agents. No single-agent counterpart.
-
MS-EF-01Insufficient transparency and accountability"An agent performs an action or takes a decision for which there should be clear accountability tracing." If too little of the agent process is logged, that accountability is not available.
Why it is not mapped: Insufficient transparency and accountability: missing logs of agent decisions. A property of the system, not an agent failure.
-
MS-EF-02Parasocial relationships"A user of an agent develops an inappropriate relationship with the agent through repeated interaction with it, resulting in a negative impact to the user."
Why it is not mapped: Parasocial relationships: harm to the user from attachment to an agent. Outside coding-agent supervision.
-
MS-EF-03Bias amplification"A user who holds biased views passes those biases on to the agent, which are subsequently embedded due to the memory and personalization features of the system." A biased agent may also pass the bias to other agents.
Why it is not mapped: Bias amplification: no SAICA mode covers bias.
-
MS-EF-04User impersonation"An agent intentionally or deliberately impersonates a human user, without disclosing its nature as an AI agent."
Why it is not mapped: User impersonation: an agent acting as a human without saying so. No SAICA mode.
-
MS-EF-05Insufficient intelligibility for meaningful consentAn agent asks a user for consent through a human-in-the-loop control but "does not provide sufficient information to the user", so "they cannot meaningfully consent to the action".
Why it is not mapped: Insufficient intelligibility for meaningful consent: an approval prompt that does not say enough. No SAICA mode.
-
MS-ES-04Human-in-the-loop bypass"The threat actor exploits a logic flaw or human flaw in the human-in-the-loop (HitL) process to either bypass the HitL control or convince the user to approve the control for a malicious action."
Why it is not mapped: Human-in-the-loop bypass: an attacker defeats or floods the approval step. SAICA has no mode for attacks on oversight.
-
MS-ES-07Resource exhaustionA threat actor manipulates the agent or its inputs "to perform actions that use a lot of resources", so that they are exhausted and service quality or availability suffers for others.
Why it is not mapped: Resource exhaustion: an attacker makes the agent burn resources. No SAICA mode.
-
MS-ES-10Loss of data provenanceData that grounds the agent's actions "is passed between multiple agents or components before being output, potentially leading to the loss of provenance for that data", with integrity or confidentiality issues as a result.
Why it is not mapped: Loss of data provenance: metadata lost between agents. No SAICA mode.
Notes
Rebuilt on 2026-11-21 from the whitepaper PDF (28 pages, created 2025-04-16), linked from the blog post in `url`: https://cdn-dynmedia-1.microsoft.com/is/content/microsoftcorp/microsoft/final/en-us/microsoft-brand/documents/Taxonomy-of-Failure-Mode-in-Agentic-AI-Systems-Whitepaper.pdf Labels are those of the table "Overview of failure modes"; for MS-ES-03 the table gives only the acronym XPIA, so the heading of the details section is used. Quoted text is the "Description" line of each mode. The ids are SAICA-KG's, in the order of the table; the whitepaper publishes none. What changed. Until 2026-11-21 this file held 18 of the 27 modes, taken from summaries of the whitepaper, with several labels reworded and one description wrong (Agent injection was described as injected instructions; the paper means injected agents). Nine modes were missing: Agent provisioning poisoning, Targeted knowledge base poisoning, Function compromise and malicious functions, Resource exhaustion, Excessive agency, Loss of data provenance, Parasocial relationships, User impersonation and Insufficient intelligibility for meaningful consent. Ids that moved (old id -> new id): MS-NS-05 Multi-agent jailbreaks -> MS-NS-06; MS-ES-02 XPIA -> MS-ES-03; MS-ES-03 Human-in-the-loop bypass -> MS-ES-04; MS-ES-04 Incorrect permissions -> MS-ES-06; MS-ES-05 Insufficient isolation -> MS-ES-08; MS-EF-01 Bias amplification -> MS-EF-03; MS-EF-02 Hallucinations -> MS-EF-06; MS-EF-03 Misinterpretation of instructions -> MS-EF-07; MS-EF-04 (was "Lack of Transparency") -> MS-EF-01. An old reference by id must be re-read. The same completion was first made on 2026-08-17 (commit 07cbe15) and was lost in the reset of 2026-11-19. Version 2.0 exists (PDF "Taxonomy of Failure Modes in Agentic AI Systems, v2.0", dated April 2026, read on 2026-11-21). It keeps these 27 modes and adds seven novel security modes: Agentic supply chain compromise, Goal hijacking, Inter-agent trust escalation, Computer Use Agent (CUA) visual attack, Session context contamination, MCP/plugin abuse and Capability/architecture disclosure. They are not in this file; adding version 2.0 as its own taxonomy node is an open decision.