A belief is not a control — 76% could stop the agent, 33% have the switch, and the clock that binds is not the AI Act

Harness's State of Agent DLC 2026, published 10 September 2026 from 700 technology professionals at large enterprises in five countries, all already running agents in production or active proofs of concept, reports the same shape in five places: 76 per cent believe they could disable a misbehaving agent inside fifteen minutes against 33 per cent with an instant kill switch, 74 against 19 on an automated release gate, 77 against 44 on inventory versus active discovery, 75 per cent confident of end-to-end security of whom 88 per cent had an incident anyway, and 58 per cent reporting more production incidents per hundred changes since agents went in. The obvious reading is that the AI Act will force the kill switch, and it is wrong twice: Article 14(4)(e) is a design duty on providers of high-risk systems, and the Digital Omnibus, Regulation (EU) 2026/1744, in force 27 July 2026, moved Annex III to 2 December 2027 and Annex I to 2 August 2028 - so for most agent stacks it is neither live nor applicable. The instrument already binding in Germany is national: the revised BSIG since 6 December 2025, where § 30 already requires incident handling and business continuity and § 38 makes the management body approve those measures, arrange and oversee implementation and attend training personally, with liability to the entity under the corporate-law rules, across a population the BSI puts at roughly 29,500 companies. The practitioner move is to convert the belief into a measurement: kill one production agent on a stopwatch, log who pulled it and how long it took, and treat that log as the § 38 oversight record rather than as documentation.

Post

Seventy-six percent of enterprise teams say they could shut down a misbehaving AI agent in under fifteen minutes. One in three has a kill switch.

Harness published The State of Agent DLC 2026 on 10 September — 700 technology professionals at large enterprises in the US, UK, France, Germany and India, fielded by Sapio Research in July, every one of them already running agents in production or active POCs.

The gaps are all the same shape. Belief on one side, mechanism on the other.

  1. 76% believe they could disable an agent inside 15 minutes. 33% have an instant kill switch.
  2. 74% trust their testing to catch failures. 19% have an automated gate that can block a release.
  3. 77% claim a complete agent inventory. 44% run active discovery tooling.
  4. 75% believe their agents are secure end to end. 88% of that group had a security incident anyway.
  5. 58% report more production incidents per 100 changes since agents went in.

Before you file this under "the AI Act will make us fix it": the stop button is Article 14(4)(e), it binds providers of high-risk systems, and the Digital Omnibus — Regulation (EU) 2026/1744, in force 27 July — moved Annex III to 2 December 2027 and Annex I to 2 August 2028. For most agent stacks it is neither live nor applicable.

The reasonable conclusion is that nothing binds you yet. It does not survive a German calendar. NIS2 has been transposed here since 6 December 2025. § 30 BSIG already requires incident handling and business continuity. § 38 makes the management body approve those measures, arrange their implementation, oversee it, and sit through training personally — with liability to the company under the ordinary corporate-law rules. The BSI puts roughly 29,500 German companies in scope.

My Monday: pick one agent in production and kill it on a stopwatch. Log who pulled it and how long it actually took. That log is not a nice-to-have. It is the only evidence that the § 38 oversight ever happened.

The gap here is evidence, not capability. Everyone in that 76% is sincere.

How long would yours actually take?

#EUAIAct #NIS2 #AgenticAI #DevSecOps #AIGovernance

First comment

Sources, in the order the argument uses them.

Harness's own release, 10 September 2026 — the survey, the methodology and every percentage in the post: https://www.prnewswire.com/news-releases/new-harness-report-reveals-enterprise-confidence-in-ai-agents-isnt-backed-by-real-controls-302875476.html

IT Pro's write-up, 11 September — independent confirmation of the same figures, plus the Harness quotes: https://www.itpro.com/software/development/agents-have-hit-the-mainstream-in-software-engineering-but-security-and-governance-practices-arent-evolving-fast-enough

AI Act Article 14 — human oversight, and the stop button at 4(e). Note what it applies to: https://artificialintelligenceact.eu/article/14/

DLA Piper on the German NIS2 transposition — in force 6 December 2025, BSI registration portal from 6 January 2026, roughly 29,500 entities in scope: https://www.dlapiper.com/en/insights/publications/2026/02/nis-2-directive-transposed-in-germany

§ 38 BSIG on the management body — billigen, Umsetzung veranlassen, überwachen, and the training duty in Absatz 3: https://www.althammer-kill.de/aktuelles/news/detail/nis-2-geschaeftsleitung

The BSI's own training material for management bodies under NIS2: https://www.bsi.bund.de/SharedDocs/Downloads/DE/BSI/NIS-2/nis-2-geschaeftsleitungsschulung.pdf

Cut for length: the same report found 42% push prompt edits through code pipelines unchanged. A prompt edit is a config change with the blast radius of a deploy, and most teams are still treating it as a string.

Claims

Sources

Corrections

None.

AI disclosure

The video for this edition contains AI-generated media: the narration is a synthetic voice, the cover art and any background imagery are generated, and the brand ident animation is generated. The script, the claims and the source verification are human-authored and human-checked. Assets carry IPTC DigitalSourceType and C2PA Content Credentials.