Salt Security reports that 17.5% of its respondents can automatically stop a misbehaving agent and 12.5% can consistently trace a complete agent path. Answer five yes or no questions and see where your organisation sits beside those figures and the Cloud Security Alliance study (31% formally adopted a policy, 54% report 1 to 100 unsanctioned agents). The surveys are shown separately, with their samples, dates and sponsors.
Survey figures come from two separate studies of self-selected respondents, both run or funded by security vendors. They show where your answers sit beside those respondents. They are not a benchmark of all organisations and not an incident rate.
0 of 5 answered
We have at least one AI agent in production.
Salt Security survey86.4%
86.4% of respondents reported at least one AI agent operating in their organisation.
Exact wording and caveat
Salt’s wording is "operating in their organization", which may be broader than "in production". This shows adoption, not good or bad practice.
We can automatically stop an agent that starts behaving unexpectedly.
Salt Security survey17.5%
17.5% said they can automatically stop an agent that begins behaving unexpectedly in production.
Exact wording and caveat
Salt: "Only 17.5% can automatically stop an agent that begins behaving unexpectedly in production." The release does not break down the other respondents, so we make no claim about them.
We can consistently trace the complete path of an agent across all our agents.
Salt Security survey12.5%
12.5% said they can consistently trace a complete agent path across all agents.
Exact wording and caveat
Salt: "Only 12.5% can consistently trace a complete agent path across all agents." The release does not break down the other respondents, so we make no claim about them.
We have formally adopted agent governance policies.
Cloud Security Alliance study31%
31% said they have formally adopted a policy.
Exact wording and caveat
CSA: "only 31% have formally adopted a policy". Separately, 50% reported at least partially documented governance policies. "Formally adopted" is the stricter measure. At least partially documented is a different, looser question.
We have unsanctioned AI agents in use (at least one we did not approve).
Cloud Security Alliance study54%
54% reported between 1 and 100 unsanctioned AI agents.
Exact wording and caveat
CSA: "More than half (54%) of organizations report between 1–100 unsanctioned AI agents". CSA’s figure counts organisations reporting 1 to 100 unsanctioned agents. It says nothing here about those with none or more than 100. Salt reports unauthorised MCP servers, not agents, so we do not use it for this question.
Salt Security: State of Agentic AI and API Security, 2H 2026
Published:
13 October 2026
Fieldwork:
Not stated on the release page.
Sample:
Over 380 security and technology professionals
Run by:
Salt Security, a vendor of API security products. Combined with Salt Labs telemetry that this tool does not use.
Funding:
Published by Salt Security; no outside funder named on the page.
Self-reported answers from a vendor’s survey sample. It is not a random sample of all organisations.
49.8%: Confirmed or suspected an AI agent took an unintended action. Respondents who "confirmed or suspected" the event. This mixes confirmed and suspected cases.
Each was claimed in secondary text but could not be confirmed on the primary page on 4 November 2026.
Salt: survey fieldwork dates. The release page gives only the publication date.
Salt: a share of respondents with unsanctioned agents. The release reports unauthorised, shadow or misconfigured MCP servers instead, which is a different thing.
Salt: a share with formal governance policies. Not mentioned on the release page.
Answers stay in this browser (local storage). Nothing is sent to PlayCISO.
TL;DR
At a glance
4 points+
Two surveys, never combined: Salt Security (over 380 respondents, published 13 October 2026) and the Cloud Security Alliance (445 responses, September and November 2025, Zenity-funded).
Both are self-reported by vendor-linked samples. They do not give an incident rate or a benchmark of all organisations.
Only figures confirmed on the primary releases are used; unconfirmed ones are omitted and listed.
Runs in your browser; answers stay in local storage. Markdown and PDF export.
How to read the comparison
Each bar is the share of that survey’s respondents who gave the answer, in the survey’s own wording. The wording differs from ours in small ways (Salt asks about agents “operating in their organization”, not strictly production), so treat the comparison as context for a conversation, not a grade.
Both studies were published by or paid for by security vendors with a product in this area. That does not make the numbers wrong, but it is a reason to read them as one view and to check your own evidence with the tests and scorecard linked above.
Frequently asked questions
What does this benchmark compare?+
Five yes or no statements about your AI agents, set beside figures from two published surveys: Salt Security (86.4% have at least one agent, 17.5% can automatically stop one, 12.5% can consistently trace a complete agent path) and the Cloud Security Alliance (31% have formally adopted an agent policy, 54% report 1 to 100 unsanctioned agents). Each is shown separately with its sample and sponsor.
Can I add the two surveys together?+
No. They have different samples (over 380 versus 445 respondents), different dates, different questions and different sponsors. We never combine them and never turn them into one average.
Do these figures show how many organisations have had agent incidents?+
No. Both surveys are self-reported by the people who chose to respond. The CSA study found 47% of its respondents had an agent-related security incident in the past year, and Salt found 49.8% confirmed or suspected an unintended agent action, but neither is a rate for all organisations, and Salt’s figure includes suspected cases.
Who ran and paid for the surveys?+
Salt Security, an API security vendor, published its own survey (release dated 13 October 2026; fieldwork dates are not stated). The CSA study was published on 16 April 2026 from an online survey run in September and November 2025; CSA analysed the data and Zenity, an agent security vendor, financed it and co-developed the questionnaire.
Some figures from the commentary around these surveys are missing. Why?+
We only use numbers we confirmed on the primary press releases on 4 November 2026. Salt’s release does not state fieldwork dates and does not report a share with formal governance policies or with unsanctioned agents (it reports unauthorised MCP servers instead), so we leave those out and list them on the page.
Are my answers stored or sent anywhere?+
No. They stay in this browser’s local storage. Markdown and PDF exports are built in your browser.
Sources
Pages below were fetched on 4 November 2026. We restate them in our own words.
445 responses, online survey run in September and November 2025. CSA analysed the data; Zenity financed the project and co-developed the questionnaire. Self-reported.
We tried to read Replit’s “Agent Habitat Architecture” post on 4 November 2026 and the page returned an error, so nothing here relies on it. Replit sells a managed agent product; treat any vendor framework as a vendor’s view, not a standard.