OWASP 2026: misuse of an AI agent’s identity and access is a top threat.
OWASP recommends a distinct identity for each AI agent, and a person’s approval before high-privilege or irreversible actions.
9 June 2026
rise in inquiries to Gartner about multiagent systems, early 2024 to mid-2025
non-human identities for every person, by Obsidian Security’s estimate
Gartner, 2025 · Obsidian Security, 2026
AI agents plan and call tools across systems without a person checking each step. Gartner reports a 1,445% rise in inquiries about multiagent systems from the first quarter of 2024 to the second quarter of 2025, and projects that over 40% of agentic-AI projects will be cancelled by the end of 2027, citing cost, unclear value and weak risk controls.
The OWASP Top 10 for Agentic Applications, published in December 2025, lists tool misuse and identity and privilege abuse among the ten highest-impact threats to AI agents. It says that “without a distinct, governed identity of its own, an agent operates in an attribution gap that makes enforcing true least privilege impossible,” and recommends per-agent identities, short-lived credentials and human approval for high-privilege or irreversible actions.
Obsidian Security, an identity security vendor, writes that non-human identities outnumber people 25 to 50 times in modern enterprises.
The EU AI Act requires high-risk AI systems to be designed so that people can effectively oversee them while they are in use (Article 14). It also gives a person affected by certain decisions made with high-risk AI systems, where the decision has legal or similarly significant effects on them, a right to a clear explanation of the AI system’s role (Article 86).
What Mandate does
Each agent gets its own sign-in, and Mandate checks each action it takes through a tool against your rules before it runs. An action your rules flag is refused, or held until a person you name approves it.
Every action goes on a signed record under the agent’s name, and anyone can check that record on their own computer.
Sources
-
OWASP Top 10 for Agentic Applications for 2026
The OWASP GenAI Security Project’s December 2025 list of the ten highest-impact threats to AI agents, including tool misuse and identity and privilege abuse, with the quote and the recommended controls above.
-
EU AI Act (Regulation (EU) 2024/1689), Articles 14 and 86
Article 14 requires high-risk AI systems to be designed for effective human oversight. Article 86 gives a person affected by certain decisions made with high-risk AI systems listed in Annex III a right to an explanation of the system’s role.
-
Obsidian Security: What Are Non-Human Identities? (February 2026)
Source for the ratio of 25 to 50 non-human identities for every person, as the vendor states it.
-
Gartner: Multiagent Systems in Enterprise AI
Source for the 1,445% rise in inquiries about multiagent systems, first quarter of 2024 to second quarter of 2025.
-
Gartner: over 40% of agentic-AI projects will be cancelled by the end of 2027
The cancellation forecast, with the cost, value and risk-control reasons Gartner gives.
More reading notes
-
CB Financial Services 2026: an SEC filing over customer data put into an AI app the bank had not approved
19 October 2026CB Financial Services told the SEC that customer names, Social Security numbers and dates of birth were handled with an AI application the bank had not authorized.
-
EY 2026: 47% say their AI governance process was skipped for urgent deployments
18 October 202698% of the 202 US senior AI executives EY surveyed say their company has formal AI governance policies in place, and 47% say the process was skipped for urgent deployments.
-
NTT DATA 2026: what the companies furthest ahead with AI prioritize in governance
9 June 2026Of the companies furthest ahead with AI in NTT DATA’s survey of 2,567 executives, 59.4% named data privacy or sovereignty across countries a top governance concern, against 49.6% of those furthest behind.
-
University of Queensland and MIT 2026: which sectors are most exposed to AI risk
8 June 2026A panel of 272 experts judged information, finance and national security the sectors most exposed to AI risk, and the authors write that the standards and enforcement AI needs are still forming or missing.
-
Schrems II and the CLOUD Act: which law reaches your data
18 May 2026Under the US CLOUD Act, a US court can order a provider under US law to hand over data it controls, wherever that data is stored.
-
IBM 2026: 18% of Canadian organizations say they have systems to govern AI across everyday operations
8 May 2026In an IBM Institute for Business Value study, 18% of Canadian organizations said they have systems to coordinate and govern AI across everyday operations.
-
United States v. Heppner: documents a defendant made with an AI tool, not at a lawyer’s direction, were not privileged
30 April 2026A US federal court found that documents a defendant made with an AI tool, not at a lawyer’s direction, were not privileged.
A free AI acceptable use policy template, written for Canadian privacy law and ready for your lawyer to adapt wherever you operate. Get the template