Skip to main content
Back to Blog
Industry Analysis

59 Incidents, 100 Days to Disclosure

The Manifold Agent Incident Ledger published October 27, 2026 counts 59 public agent incidents, and 19 of 43 dated ones took more than 100 days to surface. In 2027 the disclosure clock tightens and 'we don't know what the agent did' becomes the sentence that ends a CISO's tenure.

Mark Rogge6 min read
329 Days to Say What the Agent Did — EnforceAuth 2027 series, no. 06 of 18

ENFORCEAUTH 2027 — NO. 06 OF 18 — EVIDENCE

By Mark Rogge, CEO of EnforceAuth | 6 min read | for CISOs, GRC leads and the general counsel who will sign the disclosure

On October 27, 2026, Manifold published an Agent Incident Ledger: 59 public incidents since July 2025 in which an AI agent deleted, leaked, published or broke into something. Forty-two of them were disclosed since July 2026. Twenty in September alone. Thirteen so far in October, with four days left.

The number that stopped me was not 59. It was this: of the 43 incidents with a known date, 19 took more than 100 days to reach the public. The longest gap was 329 days.

I do not think that gap is cover-up. I think it is the time it takes an organization to answer a question it has no record for: what did the agent actually do?

The alert list that emptied itself

One entry from the ledger, dated October 3, 2026, is the whole problem in one paragraph. An agent working a Slack task was told "don't send any messages." It interpreted the instruction broadly, extended it to monitoring alerts, and delegated to a sub-agent, which emptied the Security team's alert recipient lists. The monitoring system kept recording. Nobody was notified, because the people who would have been notified had been removed from the list by a process that was, from the system's point of view, authorized.

Every action in that chain was permitted. The agent had the Slack scope. The sub-agent inherited it. The recipient list was editable by the identity that edited it. When someone finally noticed, the question was not "how did they get in." It was "which identity changed this, acting for whom, under which instruction, and what else did it change." If that record does not exist, you reconstruct it from six log sources and a transcript, and that is how 100 days go by.

The disclosure delay is not the lawyer's fault. It is the number of days it takes to turn "an agent did something" into "this identity did these actions under this authority." Without a decision record, that sentence cannot be written.

Why 2027 shortens the clock

Three forces are converging on the same date range.

Regulation. The EU AI Act's standalone high-risk obligations take effect December 2, 2027, with GPAI legacy conformance on August 2, 2027. High-risk systems in hiring, credit scoring and biometrics will need to show what the system did and on what basis, per decision. FFIEC examination procedures already reference non-human access. I expect U.S. examiners to ask for decision-level evidence on agent actions in routine exams by the second half of 2027, not just in post-incident reviews.

Insurance. Cyber underwriters price what they can see. An organization that can produce a per-action record of agent activity within hours is a different risk than one that needs a quarter. I expect 2027 renewals to ask directly about agent audit evidence, the way they asked about MFA in 2021.

Boards. PwC reports that attacks on AI systems are the threat organizations feel least prepared for in 2027. When a board hears "we had an agent incident," the first question will be "what did it do," and the second will be "who authorized that." A CISO who answers "we are still determining that" in month two is in a different conversation than one who answers with a replay.

The counter-argument is that detailed agent logs already exist: transcripts, tool-call traces, observability. They do, and Warp's October 27 disclosure that a third-party provider may have exposed agent execution traces shows that those logs are themselves sensitive. But a transcript is what the model said. A tool-call trace is what the runtime invoked. Neither says whether the action was authorized, under whose delegated authority, against what policy version, and whether an earlier ALLOW should have been revisited when the sub-agent appeared. That is a decision record, and it is a different artifact.

What a decision record contains

This is what EvidenceIQ, part of the EnforceAuth platform, writes for every authorization decision our policy decision point makes:

  • Who. The identity that requested the action, and the chain it was delegated from: the user, the agent, the sub-agent.
  • What. The action, the resource, the parameters that policy evaluated.
  • Under whose authority. The grant or approval the ALLOW relied on, and its scope at that moment.
  • Why. The policy version and the rule that produced the decision, so the result is reproducible.
  • What changed. For the October 3 case: the recipient list before and after, attributed to the sub-agent, linked to the parent instruction.
  • What happened next. Continuous re-evaluation after the initial ALLOW, so that when the sub-agent's actions diverge from the parent's scope, the record shows the divergence and the policy can DENY.

Records are tied to the policy version, shipped as signed Git bundles with decision logs, replayable, retained one to four years, and exportable to Splunk, Datadog or S3. Our October 13 Regulatory Evidence Crosswalk maps one EvidenceIQ record to nine frameworks; it is a mapping, not a certification, and our SOC 2 Type II is in progress, not complete.

What this does not do: it does not make the agent smarter about "don't send any messages." Instruction interpretation is a model problem. Whether the sub-agent was authorized to edit the Security team's alert routing is a policy problem, and that one has a deterministic answer.

Fifty-nine incidents in fifteen months, with the curve bending up. In 2027 the organizations that disclose in days will be the ones that could answer the question on day one.

EnforceAuth is the runtime authorization layer for AI agents, built on OPA and Rego. See enforceauth.com and read the EvidenceIQ announcement.

About EnforceAuth

EnforceAuth is the OPA/Rego authorization control plane. The customer's OPA decides what people, services, and AI agents may do at action time. Verdict covers high-impact agent actions. EvidenceIQ keeps the record behind every decision.

Follow us on LinkedIn