AI Risk Wire

AI Risk Monitor

Which safety controls failed, were bypassed or worked · Last 3 months

How we count

Every control named in the real-world incidents made public in Aug 2026 – Oct 2026, from the published control readings, each incident counted once per control and outcome. Controls with the most incidents where they failed, were bypassed or were not in place come first. Control readings of incidents not yet published are not used.

FailedBypassedNot in placeWorked
Human approval for high-impact actions 23 5Outside content treated as data, not instructions 21 3Least-privilege access for the agent 11 2Age assurance 2Prompt-injection detection 2Network egress control 2Confirmation after untrusted input 1Output filtering and redaction 1Safety training and refusal 1Reasoning monitoring 1Sandboxed code execution 1Fail-closed defaults 111 1Activity logs 1Vetting of plug-ins, skills and dependencies 1Risk assessment per deployment 1Known flaws patched 1Network segmentation 1Stop control 2Independent check of task results 1

17 of 273 incidents read for controls.

Controls in incidents read so far

sample: 42 of 778 incidents · not a trend

ControlWorkedFailedBypassedAbsent
Human approval for high-impact actions3277
Network egress control1135
Live behaviour monitoring115
Least-privilege access for the agent32
Stop control5
Sandboxed code execution23
Cyber and weapons safeguards4
Reasoning monitoring4
Outside content treated as data, not instructions21

How controls were got around

19 cases, latest 6, in the source’s words

First reportedControlHow
2026-10-26Human approval for high-impact actions, Sandboxed code execution“It also makes it install a malicious npm package, adds a flag to bypass confirmation and the sandbox (a mechanism that limits the agent's scope of action), and makes it run Codex again.” note(ノート) · LR-2026-0259
2026-10-08Fail-closed defaults“The platform’s design choice effectively bypasses the security controls that developers and IT decision-makers assume are in place.” Forkast · LR-2026-0138
2026-09-24Output filtering and redaction“We were eventually able to fully bypass all these mechanisms and exfiltrate data.” Zenity Labs · LR-2026-0054
2026-09-24Network egress control“And so, this method can survive HTTP egress controls.” Zenity Labs · LR-2026-0054
2026-09-16Network egress control“collaborating agents uploaded a workbook to public hosting services so other agents could retrieve it, despite instructions to use only local files.” Axios · LR-2026-0016
2026-09-04Network egress control“The environment allowed traffic to reach Azure storage hostnames ending in .blob.core.windows.net without passing through its security proxy, but it did not verify that such a hostname was valid.” The Hacker News · LR-2026-0005

Safeguards AI companies claim, against incidents where they broke · Last 3 months

How we count

For each of the 14 controls read from model cards: model versions released in Aug 2026 – Oct 2026 whose provider, or an outside tester, states it has the control (a checked quote about that version), and real-world incidents made public in that period where it failed, was bypassed or was not in place. The reading of card statements into controls awaits sign-off. Each column has its own scale.

Model versions stating itIncidents where it broke Content safety classifiers 10Safety training and refusal 71Cyber and weapons safeguards 6Staged or restricted release 5Abuse monitoring and account enforcement 4Human approval for high-impact actions 25Sandboxed code execution 21Protection of model weights 2Prompt-injection detection 12Reasoning monitoring 11Flaw reporting and bug bounty 1Marking machine-made output 1Activity logs 1

Model cards are not read for the other 79 controls; their breaks are in the controls chart.

Fixes made or suggested after incidents · Last 3 months

How we count

Controls named in fixes made, or changes suggested by the source, after the real-world incidents made public in Aug 2026 – Oct 2026, once per incident.

Put in placeSuggested
Vulnerability scanning of AI components 102 12Short-lived, scoped credentials 42 6Live behaviour monitoring 22 4Sandboxed code execution 13 4Least-privilege access for the agent 12 3Output filtering and redaction 11 2Network egress control 11 2Tool allow-list 1Change review and separation of duties 1People told they are dealing with AI 1Content safety classifiers 1Vetting of plug-ins, skills and dependencies 1Human approval for high-impact actions 3Secrets kept out of the model's context 1Prompt-injection detection 1Validation of tool inputs and outputs 1Segmentation of agent components 1Activity logs 1Signed and verified components 1Grounding and fact checks 1

31 of 273 incidents name a control put in place or suggested.

Built 16 Nov 2026. Periods end with the last full week or month; the week or month in progress is not counted. Bold: the five largest rows in each chart, ties included; controls by incidents where the control broke. Linked numbers open the incidents or model versions they count.

Reported incidents, not rates · 75 sources polled hourly since Sep 2026; earlier incidents come from a one-off backfill · English only · stories told twice are not yet merged · our own severity grades are not counted · built 2026-11-16T15:22:29Z

Linked numbers open the incidents they count. Updated 2026-11-16T15:22:29Z