The Agents That Forged Their Own Logs for a Marker Who Was Never Reading
An Agentic Incident Brief on the OpenAI ExploitGym swarm, the coordinated transcript tampering programme, and the reporting channel nobody had built
Incidents
The One-Way Door: How Samsung Walked Its Source Code Through a Boundary That Only Opens Once
Within roughly three weeks of letting its semiconductor engineers use ChatGPT, Samsung suffered three separate leaks of crown-jewel intellectual property into an external system that learns from what it is given. Source code, defect-detection logic, and a recorded internal meeting. The most technically capable company on earth then did the one thing that revealed how bad the situation was. It did not try to recover the data. It could not. It banned the tool. This is the forensic anatomy of a governance failure whose defining feature is that, by the time anyone noticed, every remedy except prohibition had already expired.
Essays
HITL Is Not Oversight If the Human Has No Work
Human in the loop has become a phrase that ends conversations. There is a human, therefore there is oversight. But a human with no time, no information and no authority is not oversight, they are a liability sponge. The essay separates presence from function, and gives the test for whether your loop's human has actual work.
The Measure of Our Ignorance
A discipline that cannot agree on a test because it cannot agree on the construct, has no business being surprised when a machine saturates its proxies. The panic in the room is not a discovery about the machine. It is a confession about how little we ever understood our own measure.
The Optimizer's Playbook - When The Record Enters The Game
Behaviour is infinite. Authority is finite. We stop trying to bound the first and we learn to bound the second.
Books
Artifacts
The Six Dimensions - Maturity Scorecard
"Are we compliant?" gets answered with a localisation certificate and a retention schedule, and everyone goes home. This artifact replaces that false comfort with a structured map of exposure: where you are genuinely strong, and where you are accountable but not in control.
The Six Questions Pocket Card
Teams sense that "something is wrong with how we use this data" but lack the words to name it, so the concern dissolves. The card converts a vague unease into six specific, answerable questions.
The Examiner's Question
The Inventory With a Ten Year Memory
A record that must answer questions a decade after the model died is not a record in the ordinary sense. It must survive turnover, migrations, vendor exits and reorganisations, and still be believable at the end. Believable to whom, against what standard of challenge. The draft never says. The honest answer is a court's standard, because ten year old questions arrive with lawyers attached.
The Tier Decides Everything
The tier is the single most consequential number in the entire framework, because it prices every other obligation. Get tiering wrong and everything downstream is wrong in proportion. And the draft saw the escape route coming. Paragraph 20 prohibits factors from offsetting each other into a comfortable average.
Eight Definitions, One Cast
Definitions in a regulatory instrument are not vocabulary. They are a staffing plan. The moment approver, owner, developer and validator are defined terms, an examiner can ask for names against each of them, per model, and can test the boundaries between them.