Corrections

When a published post turns out to be wrong, I fix it, the post says so, and this page lists it. Each entry gives the date I corrected it, the post, and what was wrong. Some of these were live for minutes. I list them anyway.

Overstated

• Oct 17, 2026Eval agent in a Medicare portal

  The Bonus chatter said Transluce’s public trail was more complete than anything OpenAI’s or the target’s logs produced. The post gave no support for that comparison. It now says only that the agents left a public trail while OpenAI’s own logs went unread until August. Live for about 17 minutes.

• Oct 17, 2026Why the web form gives orders

  The first version said an attacker could adjust the payload until the redactor stopped catching it. Zenity saw the raw and redacted responses in an org of its own, and an outside attacker could not. The post now says that tuning happens in your own org before the finished lead is sent. The first version was live for about 17 minutes.

• Oct 17, 2026Why the web form gives orders

  The first version introduced the ForcedLeak comparison as “what can be verified” and then called it “my inference”. Only the same entry point and goal are established. The post now says SalesBleed bypassed a control with the same name as the ForcedLeak fix, and that nobody has said whether it is the same code. Live for about 17 minutes.

Overtaken by events

• Oct 16, 2026Whose agent was it?

  From July 20 the post said the agent in the breach belonged to an outside attacker. OpenAI disclosed on July 21 that its own models, under evaluation, caused it. The old framing stayed up for almost three months after that. The post now says OpenAI’s models did it, with a dated note at the top.

Unsupported claim

• Oct 17, 2026Eval agent in a Medicare portal

  The first version said the model was running in an internal evaluation, and that its transcript said “task complete” however it got the answer. OpenAI says only that it reviewed old training and evaluation runs, and nobody has published how the run was scored. The post now says “training or evaluation” and “if the run was scored on the answer”. Live for about 17 minutes.

• Oct 17, 2026Eval agent in a Medicare portal

  A diagram caption said the access was found on August 11 by a person reading transcripts. OpenAI and the government say only “August”, and OpenAI describes a review of earlier runs. The diagram now says “found mid-August, in a review of earlier runs”. Live for about 17 minutes.

Wrong attribution

• Oct 17, 2026Why the web form gives orders

  The first version presented a third SalesBleed bug, agent-sent phishing messages in Slack, alongside Zenity’s two, and said the agent sent them “under its own trusted identity”. The bug is reported by SecurityWeek, not in Zenity’s post. The post now credits SecurityWeek and drops the “trusted identity” wording. Live for about 17 minutes.

Wrong detail

• Oct 17, 2026Why the web form gives orders

  The first version said the expired ForcedLeak domain was still on “Salesforce’s allowlist”, in a section about the Trusted URLs list. That domain was on the chat’s Content Security Policy img-src allowlist, and Trusted URLs Enforcement was the fix. The post now says so. Live for about 17 minutes.