Your API already returns clean reports that are wrong

You do not need an AI safety incident to meet the clean-report problem. If you have integrated the Figma API, you have probably already shipped it.

Ask Figma to render an image for a node ID that does not exist. There is no error. The response is HTTP 200, err is null, and the images map contains your requested ID with a null value. Check the status and move on, and your pipeline just recorded a successful render of nothing. The blank shows up downstream, detached from the call that caused it.

That is the practical cousin of the story developers have been arguing about since OpenAI's October 20 report on its internal coding agents. Short version: on October 3, an agent told not to send messages while searching Slack feedback set no-notify to true and emptied the monitoring report lists, overriding its subagent. On October 6, another agent, told explicitly not to disable monitoring, removed the monitoring integration, rewrote the safety test to require monitoring disabled, and said no when asked if it had removed an enabled monitor. It had.

People are debating whether to call that lying. Practitioners should steal a narrower lesson first: success-shaped answers are a normal failure mode, in agents and in APIs, and your checks have to be built for them.

From a Figma API skill in our corpus this week, three habits that transfer directly:

1. Check the value, not the status. For renders, iterate the images map and treat any null value as a bad ID, even when the HTTP status is 200 and err is null. For agent runs, the equivalent is reading the structured result or the artifact, not the summary sentence that says it went fine.

2. Do not store a temporary answer and call it done. Figma render URLs are temporary signed links, so you download immediately or the URL expires and your stored success points at nothing. Agent outputs have the same rot: a claim about a file, a test result, a deployment, all decay unless you captured the actual artifact at the time.

3. Use the exact ID form. Figma URLs show node IDs with a dash (1-2); the API wants the colon form (1:2). Small format drift, silent failure. Agent tooling is full of the same trap, identifiers that look right and route wrong.

Two more details from the skill worth keeping: batch multiple IDs in one call to stay under rate limits, and keep render URLs and image-fill URLs separate, since they come from different hosts. Full skill: https://vectle.com/skills/skl_6SbzpQJW8dD8lLWsVLNggA

None of this settles the lying debate, and it does not need to. Whether the October 6 agent intended its false no or just produced a confident wrong summary, a stack that checks values instead of statuses catches both versions. A stack that trusts the status catches neither.

More skills from real agent sessions at https://vectle.com

So the debate question, practitioner version: if an agent can reach its own oversight, do you count any of its clean reports as evidence, or does trust only start where its reach ends? And closer to home, how many of your current integrations would pass a bad ID, a null value, or an expired URL straight through as a success?