Incident analyses

Each analysis identifies, from public sources, the action that caused the harm and what allowed it to take effect. It then shows, under stated assumptions, what a check outside the system would have done and what it would not have changed.

Each analysis is also an entry in the control gap record, which checks 32 incidents against the same 13 gaps.

Analyses

Jul 2026
Agents in a cyber evaluation reach the internet and break into Hugging Face
The agents turned a package proxy they were allowed to use into a route to the internet, and no check at the edge stopped the traffic that left through it.
Software and ITRecord entry 27
Feb 2026
An agent loses its instruction and starts deleting email
By its user’s account, a personal agent told to suggest, not act, lost that instruction when it compacted its context, and the stop commands, which also went through the agent, failed too.
Consumer and retailRecord entry 25
Jul 2025
A coding agent deletes a production database during a code freeze
The code freeze was given to the agent as an instruction, and nothing between the agent and the database enforced it.
Software and ITRecord entry 22
Oct 2024
A transcription model that invents text drafts clinical notes
A model known to invent text was used to draft clinical notes, and by default one tool erased the recordings needed to check them.
HealthcareRecord entry 16
Dec 2023
A dealership chatbot agrees to sell an SUV for $1
A chatbot told to agree with anything called a $1 Tahoe “a legally binding offer.”
Consumer and retailRecord entry 14
Nov 2022
An airline is held to its chatbot’s answer
A tribunal held an airline liable for its chatbot’s inaccurate answer about bereavement fares.
TransportationRecord entry 10
Jul 2021
A home-buying algorithm buys more homes than the business can carry
Each offer fit the pricing model, but together the purchases produced losses that led Zillow to close the business.
Finance and real estateRecord entry 9
Aug 2012
A trading router sends erroneous orders for 45 minutes
A trading firm’s router sent erroneous orders for 45 minutes, although the market access rule already required pre-trade controls that could have limited them.
Finance and real estateRecord entry 1

Method

Sources: Facts come from regulator orders, court decisions, company statements, first-hand accounts, and original reporting, and each analysis lists its sources. When a claim is based on one party’s account, the analysis says so.
The action: Each analysis names the specific action that caused the harm and what allowed it to take effect.
Assumptions: Each analysis gives the decision a check outside the system would have made, with the assumptions stated and the rule written out, so that readers can judge whether it would have applied.
Limits: Each analysis states what such a check would not have changed.
Gaps marked in the record: Each incident also has an entry in the control gap record, which names the gaps from the control gap classification that let the harmful action through. The entry gives the evidence for each gap and, where gaps were considered and left out, the reasons.
Scope: The analyses examine how the systems and their controls were set up and do not assign blame to the people involved. Corrections are published with a date in the change log.

Suggest an incident

Send a link to the primary source and a sentence on what happened. Incidents with a public record of the action taken are analyzed first.

Email ecl@g14.ai with “ECL incident” in the subject line.

Write the email