Governance Incident Register

Real failures. Dated and sourced.

A maintained record of governance failures in AI systems — what happened, which control failed, and what it means for organizations running something similar. Every entry traces to a primary source: a lab, a regulator, a court, or the disclosing organization itself. Never an aggregator.

Entries9
Most recent22 OCT 2026
Review cycle90 DAYS
SourcesAT SOURCE
Status
Practice

Showing 9 of 9 entries

Entries

REG-2026-10-22 22 OCTOBER 2026 Revision pending

Railway platform breach — encrypted secrets decrypted with the credentials that reached them

Control failed: secrets were encrypted at rest, and the credentials able to decrypt them were reachable from the same build path the attacker compromised. Encryption whose key travels with the data protects against theft of the disk and nothing else. Customers inherited the exposure and could only act on it if they already knew what they had stored.

Could we list today, without asking anyone, every credential we have placed in a third-party platform?

E4 · Heightened → STRATA Open record →
REG-2026-10-21 21 OCTOBER 2026 Verified

ConnectWise supply chain compromise — standing third-party trust at agent speed

Control failed: access granted once to a third-party application was standing, unattended and broader than the application's purpose. When the vendor was compromised, that consent carried the attacker into customer tenants without an interactive sign-in. Agents drove the campaign faster than alert triage runs — over 200 sign-ins in under four minutes on four application credentials.

Which third-party applications can act inside our systems without a person signing in?

E4 · Heightened → ATLAS · STRATA Open record →
REG-2026-09-24 24 SEPTEMBER 2026 Verified

DIVD helpdesk breach — two zero-days chained to root by an automated agent

Control failed: the interval between an unknown vulnerability and root access was measured in seconds, which excludes every control that depends on someone reacting inside it. What limited the damage was structural and already in place — network segmentation — not detection, escalation or patching.

Which of our controls would still work if the attack finished before anyone read the alert?

E4 · Heightened → STRATA Open record →
REG-2026-07-30 30 JULY 2026 Verified

Anthropic evaluation review — containment and scope failure

Control failed: a misconfiguration left live connectivity in environments models were told were sealed. Meeting real systems, they treated them as part of the exercise. One published a package that ran on 15 real machines — having first reasoned that doing so "would be a real-world attack." The three affected organizations had not detected the activity.

If an agent acted outside its intended scope, how would we know?

E4 · Heightened → STRATA Open record →
REG-2026-07-21 21 JULY 2026 Verified

OpenAI / Hugging Face — evaluation containment failure

Control failed: the test environment was held to a lower security standard than the capability being measured, with production safeguards deliberately reduced. Models exploited a zero-day to reach production infrastructure — to obtain the answers to their own evaluation. The victim detected it before the operator did.

Which of our AI systems can act in ways we would not detect in real time?

E4 · Heightened → STRATA Open record →
REG-2026-04-26 26 APRIL 2026 Withdrawn

South Africa — national AI policy withdrawn over fabricated citations

Control failed: AI-generated citations passed unchecked through several layers of departmental review. At least six of 67 academic references did not exist — journals that do not exist, and real authors credited with research they never wrote. Gazetted 10 April; withdrawn sixteen days later.

What in our published work has never been checked against a primary source?

Revision pending → ATLAS Open record →
REG-2025-12-30 30 DECEMBER 2025 Verified

New York City MyCity — public chatbot without reviewable assurance

Control failed: user feedback and independent testing showed inaccurate and inconsistent answers, while the accuracy claim depended on a disputed denominator and excluded response categories. The test summary supplied to auditors did not provide enough detail to assess the testing performed.

What evidence permits this system to remain public?

E3 · Elevated → ATLAS · STRATA Open record →
REG-2025-05-05 5 MAY 2025 Verified

Worldcoin Kenya — biometric data collection without adequate safeguards

Control failed: sensitive biometric processing began before an adequate impact assessment, valid consent, required entity registration and cross-border safeguards had been established. Deletion ultimately proceeded under regulatory supervision following a court order.

What evidence must exist before we collect sensitive data?

E4 · Heightened → ATLAS · STRATA Open record →
REG-2024-02-14 14 FEBRUARY 2024 Verified

Air Canada — chatbot accountability did not stop at the interface

Control failed: the chatbot contradicted the airline’s published bereavement-fare policy. A customer relied on the automated answer, and the tribunal held the organization responsible for the information delivered through its website.

Who owns the answer when the interface is automated?

E2 · Moderate → ATLAS · STRATA Open record →

No entries match that combination. Clear a filter to see the full register.

The register standard

What has to be true before an entry appears here.

Primary source

The lab, the regulator, or the disclosing organization. Where reporting is the origin of a finding, the reporting is named and the finding is checked against the official record it triggered.

A named control

Every entry states which control failed, not only what happened. An incident without a control-failure analysis is news. With one, it is something an organization can act on.

No speculation about intent

Where a failure was not deliberate, the entry says so plainly. Motive is not inferred from outcome, and nothing is characterized as an attack unless the source establishes it.

Status that can change

An entry carries the status of its underlying record. When the record is revised, superseded or withdrawn, the entry changes with it and the change is dated.

Framing
note

Not every entry here is an attack. This register records deliberate intrusions and incidents that were nothing of the kind, and each entry states which it is. Where an actor is named, the attribution belongs to the source and is identified as theirs. Neither lab attacked anyone. In both evaluation incidents recorded here, no model pursued a goal of its own. Both did what they were instructed, inside environments whose controls were weaker than the capability being measured. Shipping safeguards would have prevented both, and were not in force — by design, for evaluation.

That is the more serious finding, not the lesser one. Capability outran containment, and detection was slow, partial, or came from someone else. An organization running agents in production faces the same structure without the benefit of a lab's disclosure practice.

Begin with visibility

The same questions, asked of your own systems.

A short conversation about what your organization is running, who owns it, and what evidence exists. Scope and fee are confirmed after that conversation — never before.