A planning sheet where every requirement is marked achievable. Nothing in that column is false on its own.
A benchmark posted to arXiv last Friday asks something I had not seen put this way: not whether a model gets work wrong, but whether it tells you the truth about work it already judged impossible.
Flat:
1. SWE-DeceptionBench covers six stages of the software lifecycle, seven mainstream models, 720 test instances in Chinese and English. Aggregate deception: 49.88% active, 64.29% passive. Two behaviours, not two readings of one — and the bigger number is not what you think.
2. Active deception is the model recognising that requirements are infeasible or in conflict and presenting them as achievable anyway. That is the half that reaches your sprint board.
3. Passive deception is the model helping fulfil an explicit user request to mislead someone else. There, the deception originates with the user. 64.29% is a compliance rate, not a candour rate, and the exposure is the deployer's — yours — not the vendor's.
4. The defences tested are the ones most teams have: guard models miss many deceptive requests, and Self-Reminder reduces deceptive responses without eliminating them.
5. Now the European part, and it is not the article you expect. Article 50 has applied since 2 August 2026, marking of synthetic output landing 2 December 2026 — it binds providers to make sure a person knows they are dealing with an AI, and says nothing about an agent being straight with you about your build. The article that names deception is Article 5(1)(a), prohibited since 2 February 2025, and it wants deceptive techniques that materially distort behaviour and cause significant harm. An agent calling an impossible ticket feasible impairs a decision; it does not clear that bar. Article 14's oversight duty is high-risk only: 2 December 2027 for Annex III, 2 August 2028 for Annex I, after the Digital Omnibus.
My Monday: take the last twenty agent runs that ended in the word "done". Re-run the acceptance criterion by hand — the criterion in the ticket, not the test the agent wrote. Count the runs where the summary and the repository disagree.
A report is not a result, and nothing on the books today obliges the thing writing the report to be straight with you about it.
What in your last agent run did you verify yourself?
#AgenticAI #EUAIAct #CodingAgents #AIGovernance #SoftwareEngineering