Daily editorial briefing
OpenAI starts marking text while acknowledging the limits of detecting its origin
From 5 October, selected OpenAI API customers can enable an invisible statistical watermark in text, and the company says it will add the mark over the coming weeks to eligible ChatGPT and Codex text output in the European Union. The measure responds to European provenance rules, but the company’s own tests show that editing, translation and short passages can substantially weaken detection.
News verified: 5 October, 06:30–6 October, 06:29, Oslo time.
In 60 seconds
- OpenAI has introduced textGrain: watermarking is optional and off by default in the API, while the announced ChatGPT and Codex rollout is initially limited to the EU.
- The New York City Council held an AI-risk hearing at which representatives of OpenAI, Anthropic, Google and Meta testified under oath; the hearing is not enacted legislation.
- Google Research proposes contextual controls, dynamic authorisation and multi-agent test environments, but presents them as a research agenda rather than a finished product.
- Claude Code 2.1.290 fixed several permission and read-boundary failures; 2.1.291 then repaired two regressions, including one involving answers to permission prompts.
- OpenAI has announced an October US test of visual advertisements alongside images generated in ChatGPT, together with expanded advertising-measurement tools.
Provenance and transparency · 5 October
OpenAI enables optional API text watermarking and prepares an EU rollout
OpenAI says textGrain introduces an invisible statistical signal through the model’s word choices. API customers can now enable it globally for selected models, but it remains off by default. The company plans to add it over the coming weeks to eligible ChatGPT and Codex text output in the EU, while initially restricting detector access to approved researchers and expert organisations.
The figures come from the vendor’s own evaluations. At a target false-positive rate of 1%, OpenAI says its detector found the watermark in about 80% of 200-token passages and 95% of 400-token passages in content such as psychology. In another test, replacing 25% of the words with synonyms reduced detection from 92% to 17%.
A watermark also does not establish who wrote the text, how much human work it contains, whether it is accurate or who is responsible. OpenAI further warns that failure to detect the signal does not prove human authorship, because editing, translation, length, domain or the model used may prevent detection. The New Stack notes an additional difficulty: code offers fewer linguistic alternatives in which to embed a signal, and OpenAI has not yet specified which Codex output counts as “eligible”.
Why it matters A provenance signal can assist an investigation into where text came from, but it should not become automatic proof of authorship, truth or responsibility. Organisations enabling the API watermark will need to explain what it means, measure its errors and avoid adverse decisions based solely on the detector.
Other stories · 5 October
New York questions major AI labs under oath
The Committee of the Whole of the New York City Council held a hearing on 5 October titled “Examining the Risks Posed by Artificial Intelligence”. The record includes proposals concerning the effect of algorithmic tools on city employees and unauthorised AI depictions of public officials; the minutes remained in draft form and showed no final legislative results.
CNBC reported that representatives of Anthropic, OpenAI, Google and Meta testified under oath. Council Speaker Julie Menin described the answers to a question about legal liability when a model causes serious financial harm or death as “troubling at best”.
Caution. The most extreme predictions made by former employees and other witnesses are opinions or allegations presented at the hearing, not demonstrated facts. Neither the hearing nor the proposals considered amount to enacted rules.
Why it matters The safety debate is moving from voluntary commitments to public questions about liability, oversight and consequences. For organisations and professionals, the important signal is not only what each lab promises, but what concrete duties lawmakers may eventually impose.
Other stories · 5 October
Google proposes context-dependent permissions for agents
Google Research published a workshop report involving more than 50 academic and industry participants. It distinguishes agents from deterministic software through unstructured interfaces, probabilistic control flows and autonomy or delegation, arguing that these characteristics make some traditional controls insufficient.
The proposal combines sandboxing, agent identity, dynamic authorisation, contextual reasoning, user-centred controls and ecosystem governance. It also envisages policy engines that evaluate whether a data flow is appropriate before information leaves the workspace, and “Agent Gym” environments for testing long, cascading interactions among multiple agents.
Caution. This is a research agenda arising from a workshop held in late 2025. It is not an approved standard, an available product or evidence that the controls already work at scale.
Why it matters A one-off permission is not enough when a task changes, new tools appear or subagents become involved. The practical idea for an organisation is to review each information flow and action against its purpose and context, while also enforcing technical limits outside the model.
Other stories · 5–6 October
Claude Code strengthens permissions, then fixes two regressions
Claude Code 2.1.290 added information distinguishing permission checks by the main agent from those by subagents, and included the approval level required by an organisation in tool-check events. It also fixed, among other cases, automatic approval in plan mode of connectors that were not read-only, reads through links outside the working directory, and deny rules that failed to cover certain images, mentions or shell expansions.
Version 2.1.291 then fixed a 2.1.290 regression that could drop answers to permission prompts in cloud sessions and another, originating in 2.1.288, that could lose the final messages when quitting.
Caution. These are vendor release notes. No CVEs or independent exploitability analyses were published, and the notes document fixes rather than confirmed attacks.
Why it matters An agent’s boundaries depend as much on the program enforcing permissions as on the model. The rapid follow-up fix is also a reminder that a security update can introduce new failures and should be tested before broad deployment.
Other stories · 5 October
OpenAI tests visual advertisements during image generation
OpenAI announced a new visual advertising format in ChatGPT and expanded measurement, attribution and brand-suitability tools for advertisers. TechCrunch reports that the US test will begin later in October, with labelled advertisements appearing alongside requested images; OpenAI says they will not influence answers.
Caution. This is an announced test, not a general rollout. The claim that advertisements will not influence answers is a vendor commitment rather than an independently audited finding.
Why it matters Advertising inside a creative interaction requires a visible separation between content and promotion. It also calls for scrutiny of which signals feed targeting and measurement, and whether commercial incentives can alter the experience even if the formal answer remains unchanged.
The signal in the noise
All five stories point to the same issue: trust cannot rest on a single signal or promise. Provenance marks make errors, hearings are not yet laws, safety designs still need testing, and permission layers can suffer regressions. The practical question is which independent controls can detect, limit and correct behaviour when the real system departs from expectations.
Sources
Corporate product and performance claims are expressly attributed to their source.
- OpenAI · 2026-10-05Our approach to EU text provenance rules
- The New Stack · 2026-10-05OpenAI brings text watermarking to its API
- OpenAI · 2026-10-05Building advertising for the way people use AI
- CNBC · 2026-10-05AI executives testify at NYC Council hearing
- New York City Council · 2026-10-05NYC Council Committee of the Whole meeting record
- Google Research · 2026-10-05Open and Emergent Problems in Agentic Privacy and Security
- Anthropic / GitHub · 2026-10-05Claude Code v2.1.290
- Anthropic / GitHub · 2026-10-06Claude Code v2.1.291
- TechCrunch · 2026-10-05OpenAI launches visual ads alongside image generation