Salt Security published a survey on November 19 that finds executives granting AI agents authority faster than they can see what the agents do. The report, Autonomy Before Control, rests on Censuswide interviews with 500 C-suite leaders, 250 in the US and 250 in the UK, at organizations with 50 or more employees, fielded September 10 to 18.
Salt reports that 92% of respondents would allow AI agents to take at least some independent action. Among organizations that use or have explored AI, 77% have seen an AI system or agent act outside its intended or authorized behavior. Sixty-four percent have restricted, significantly delayed or canceled an AI project over security, governance or regulatory concerns. And 97% say they keep an inventory of their agents, while 19% of those admit the inventory is not kept consistently up to date.
Salt sells agentic and API security products, so the survey also supports its own product pitch, and the full report is the place to check how questions were worded. Michael Nicosia, Salt’s co-founder and COO, says in the release: “Governance cannot be a one-time approval.”
In our reading, the 97% and 19% figures belong together. The survey suggests an inventory alone is not enough, since agents are being added across separate teams. Agents inherit access through existing applications, identities and APIs, according to Salt, so a security team gets a more useful list by asking each identity provider and API gateway what an agent token can reach. That is a list built from the systems that grant access. On the attacker side, CyberTech has covered Google’s report on a framework that turns patches into exploits without a human and coverage of the FBI’s ARTEX alert.
Source: Salt Security, Autonomy Before Control: The 2026 C-Suite Report on AI Agent Risk, on PR Newswire