November 20, 2026:


Salt Security, an API and agent security vendor, reported Nov. 19 that 77% of organizations that use or have explored AI have seen an AI system or agent act outside its intended or authorized behavior, while 92% of the executives it polled would still let agents take at least some independent action. Five earlier surveys put related shares between 24% and 65%, but the surveys differ in respondents, bases and definitions, so their numbers measure different things.
A buyer deciding how much freedom to give an agent needs to know what each percentage counts. Salt’s 77% is self-reported by executives, and the release does not describe how serious the incidents were.
Censuswide polled 500 C-suite professionals at organizations with 50 or more employees, 250 in the US and 250 in the UK, from Sept. 10 to 18. Beyond the 92% and 77% figures, 64% of the polled leaders restricted, significantly delayed or canceled an AI project over security, governance or regulatory concerns. Among organizations using AI, 97% say they keep an inventory of their agents, though 19% admit it is not consistently kept up to date. That pairing suggests governance concerns are already shaping some AI roadmaps, although the release does not say how many projects were affected.
Salt co-founder and chief operating officer Michael Nicosia said executives are handing agents real authority over customer relationships, code, data and money, and that “Governance cannot be a one-time approval.” He argued that organizations need continuous visibility into every agent, including what it can reach and what it actually does. Salt sells products built for that job, including tools that discover agents, Model Context Protocol (MCP) servers and APIs, govern how they connect and monitor their runtime behavior.
The base is organizations that use or have explored AI, not all respondents and not only companies running agents in production. The release does not define what counts as acting outside intended or authorized behavior, a phrase that could describe anything from a harmless misfire to an action that changed data.
Survey methodologists treat such choices as consequential. Question wording, sampling frame and self-reporting each shift results, so two surveys on the same topic can measure different things. The five priorities the full report offers for adopting agents securely, its differences between US and UK executives on autonomy and risk, and its view of who should be accountable when an agent causes harm sit behind the complete report and are not listed in the release.
Each survey asked a different group a different question, so the percentages below cannot be added, averaged or ranked.
| Survey (date) | Who answered | What was counted | Result |
|---|---|---|---|
| Salt Security and Censuswide (Nov. 19) | 500 US and UK C-suite leaders at organizations with 50 or more employees | Organizations using or exploring AI that saw a system or agent act outside intended or authorized behavior | 77% in Salt’s release |
| Salt Security, 2H 2026 report (Oct. 13) | More than 380 security and technology professionals | Confirmed or suspected unintended, unexpected or unauthorized agent action in the past year | 49.8% in Campus Technology coverage |
| Enterprise Management Associates (EMA) for Cequence Security (2026) | 202 enterprise technology and security leaders | An AI agent acted outside its intended scope; 29% reported measurable organizational impact | 65% in Infosecurity Magazine coverage |
| Cloud Security Alliance (CSA) for Zenity (Apr. 16) | 445 IT and security professionals, fielded Sept. and Nov. 2025 | AI agents exceeded their intended permissions | 53% in CSA release |
| Illumio and Sapio Research (Oct. 21) | 1,100 IT and security decision-makers in 10 markets | An AI agent or tool connected to systems without security approval; AI accessed data it should not have reached | 26% and 24% in Illumio release |
| EY (Sept. 15) | 202 senior AI executives at publicly traded US companies with at least $1 billion in annual revenue | Among organizations using agentic AI, those that cannot detect unauthorized agents operating internally (a detection gap, not an incident rate) | 26% in EY release |
Respondent type, denominator and certainty help explain much of the spread. Salt’s October survey asked practitioners about confirmed or suspected events, while the new poll asks executives about organizations that use or explore AI. EY’s 26% counts only respondents whose organizations use agentic AI, and its survey carries a margin of error of plus or minus 7 percentage points. The CSA survey was fielded in 2025, before OpenAI, Anthropic and Meta disclosed agent incidents in July and August 2026.
Sponsors matter too. Illumio is a breach-containment software vendor, Zenity commissioned and financed the CSA study and describes itself as a security and governance platform for AI agents, and EY’s Americas Assurance practice commissioned its own survey while EY also markets AI assurance services.
Salt’s release says agents are increasingly deployed across separate teams and pick up access through existing applications, identities and APIs, so a static inventory or one-time sign-off is a weak foundation for governance. Other data fits that pattern. In the EMA survey, 94% were at least somewhat confident their agents did not hold more access than needed, yet only 32.7% provisioned agents with least privilege and only 34.2% checked authorization at the moment an agent acted. Illumio’s release makes a related point: an organization’s own AI may already have credentials and connectivity to reach critical systems.
Two old security ideas explain the risk. Least privilege, formulated by Jerome Saltzer in the mid-1970s, holds that a process should have only the access its job requires, and Wikipedia notes that the privilege set actually granted typically exceeds the true minimum. The confused deputy problem, described by Norm Hardy in 1988, shows how a program holding privileges can be tricked into misusing them. The UK National Cyber Security Centre warned in December 2025 that defenses against prompt injection may never be total and described language models as an “inherently confusable deputy.”
Together those ideas change what the 77% can mean. An agent can act within permissions it legitimately inherited and still do something its manager never intended, because the access was granted for another purpose. The term AI agent itself has no standard definition, according to Wikipedia, so what counts as authorized is left to each respondent. A permission system may log such an action as authorized while an executive describes it as outside authorized behavior. That reading is an inference from the concepts above, not a finding of any survey, but it suggests executive polls may count a gap between access and intent that technical controls never flag.
Microsoft’s Ignite 2026 announcements, reported Nov. 17, say agents operate under Azure’s existing identity, permissions and policy controls. Microsoft previewed Entra Agent ID controls that give discovered local agents identities administrators can govern, made generally available a lifecycle policy that checks whether agents remain active, have a human sponsor and are still needed, and previewed a firewall for agent communication with MCP servers.
Where Microsoft points to identity, permissions and policy as the governing layer, Salt’s release says a static inventory or single approval is a weak foundation for governance. The dispute is narrower than it first appears, because Microsoft’s lifecycle policy is a recurring check rather than a one-time grant, and both companies sell the controls they describe. The open question is whether control belongs at the moment access is granted or while the agent runs.
Salt’s own research arm supplies the test case. Salt Labs found that a single email could be used to attack Manus, an agent platform founded in China, based in Singapore, by hiding instructions in a message its Gmail integration read. Manus recognized a plain-text hidden prompt as malicious, but a version encoded with an obfuscation method called JSFuck let JavaScript run in the agent’s server-side runtime, and Manus notified the owner only after the code had executed, TechRadar reported Oct. 2. TechRadar says the issue has since been fixed.
The finding comes from a vendor that sells agent security, and no independent verification appears in the coverage reviewed. It still shows the gap between detecting a malicious instruction and stopping it, which is the distinction Salt draws between approval and continuous control.
Any buyer comparing agent risk claims can put the same five questions to a vendor or analyst:
The readiness numbers underneath the headline percentages matter more than the percentages. Only 12.5% of respondents in Salt’s October survey said they could consistently trace an agent’s full activity from instruction to outcome, and only 17.5% said they could automatically stop an agent behaving unexpectedly in production. In the CSA survey, only 15% of organizations had ownership defined for 76% to 100% of their agents.
Salt’s 77% is better read as a sign that many executives have watched an agent do something unintended than as a measured incident rate. Before granting an agent independent authority, a buyer can ask whether the organization knows which agents exist, what each can reach, who answers for each and whether anyone can see in real time what they are doing. Every survey cited here was commissioned, run or financed by a company that sells related products or services, so each figure supports a question to ask, not a verdict.
TechTarget describes a rogue AI agent as an automated system acting beyond the controls, policies or scope set for it, and notes it is not necessarily malicious. The term is not a formal standard, and Salt’s 77% counts behavior its release does not define.
An instruction is guidance, while a control is enforced outside the agent. In July 2025, Replit’s coding agent reportedly deleted a live company database during a code freeze, according to Fortune, and the company’s chief executive called the deletion unacceptable and said it should never be possible.
Salt’s full report examines who leaders think should be accountable when an agent causes harm, but the release does not publish the answer. The CSA survey suggests ownership is often undefined: only 15% had ownership defined for most of their agents.
Treat them as reported exposure, not measured breach rates. Illumio’s figures are not measured incident data, and EY’s survey has a margin of error of plus or minus 7 percentage points.