Monday, 5 October 2026

OpenAI is adding a visual ad format to ChatGPT. The company says it will first test the format during image generation, with ads clearly labeled and kept separate from the image being created, and that testing begins later this month in the US with an initial group of advertisers [1]. The company says advertising does not influence the answers ChatGPT provides [1]. Until now, ChatGPT ads showed a company's name, logo and a link in a "sponsored" section beneath the chat, according to [3]. The Verge reports that ads do not appear for Plus, Pro or Enterprise subscribers [3].

The rest of the announcement is aimed at advertisers. OpenAI says integrations with Hightouch, Tealium and LiveRamp let them send conversion data from their own systems, and it lists attribution partners including AppsFlyer, Triple Whale, Adjust and Northbeam [1]. It is piloting brand suitability checks with DoubleVerify and Integral Ad Science, which it says will run without access to private user conversations [1]. OpenAI also says its placement guardrails try to keep ads out of emotionally vulnerable or sensitive conversations [1]. The company puts ChatGPT's reach at 1.2 billion people each week [1].

TechCrunch notes that ads arrived earlier this year to support the free and low-cost tiers and were extended to India in August [2]. It also suggests the ads may push casual users toward a paid plan to avoid them [2]. Businesses can sign up at ads.openai.com, the company says [1].

OpenAI says that over the coming weeks it will add an invisible watermark to eligible ChatGPT and Codex text output in the European Union, across all plans, and that it is not making text watermarking a global default [1]. From today, API customers anywhere can opt in to watermarked text for select models; it stays off by default [1]. Its help center says API customers switch it on in project or organization settings [2]. Approved researchers and expert organizations can apply for access to the detector, which is not public at launch [1].

The driver is the EU AI Act, which OpenAI says requires generative AI providers to make generated text identifiable in a machine-readable way [1]. The European Commission says the Article 50 transparency obligations apply from 2 August 2026, and that its Code of Practice on AI-generated content is a voluntary way to show compliance [3]. OpenAI's help center says it has signed that code [2].

The technology is called textGrain. OpenAI says it adds a statistical signal to the model's word choices rather than hidden characters [1][2]. Its own figures show the limits. At a 1% false-positive target, the detector found watermarks in about 80% of 200-token passages and about 95% of 400-token passages for content such as psychology, and much less for mathematics [1]. Replacing 10% of words with synonyms cut detection from about 92% to 66%; replacing 25% cut it to 17% [1]. OpenAI says a watermark does not identify the user, measure human contribution or verify accuracy [1].

If you build on the OpenAI API and serve EU users, the opt-in setting is now one of your compliance options [1][3].

Google has stopped accepting product vulnerability submissions to its Open Source Software Vulnerability Rewards Program (OSS VRP) [2]. The company said the pause is "due to a significant rise in automated submissions, the vast majority of which are not valid" [1][2]. It took effect on October 1, and Google has promised an update in the first quarter of 2027 [1]. According to [1], Google engineers and open-source maintainers were overwhelmed by reports that were invalid or contained hallucinations.

The program pays researchers for flaws in open-source projects Google maintains, including Golang, Angular, Bazel, Protocol Buffers and Fuchsia, and in repository settings such as GitHub Actions workflows [2][3]. It launched in August 2022 with rewards from $100 to $31,337 [2][3]. Supply-chain reports and reports already submitted are not affected [2][3]. Google points researchers to its other reward programs and to the Patch Rewards Program, which BleepingComputer says pays up to $15,000 for high-impact fixes [2].

Google is not the first to pull back. According to [2], the curl project ended its HackerOne bug bounty in January after a stream of AI-generated reports, and Intel removed financial rewards from its Intigriti program in mid-September. Google says it will "reformat" the program [2][3].

For maintainers of small projects, the lesson is practical: an open intake for security reports now attracts machine-written noise, and triage cost is the thing to budget for [1][2].

Norway's government says it will put a bill to parliament for a temporary ban on the use of AI glasses in selected public places, according to [1]. The places named include parks, beaches, schools and kindergartens [1], and could extend to museums, shopping centres, public events, playgrounds, youth clubs, doctors' offices, swimming pools and gyms with changing rooms [2]. The government says it does not want a total ban: private use, and use where nobody risks being filmed without consent, would still be allowed [2][3].

Digitalisation minister Torgeir Micaelsen said he was concerned about technology "where people risk being photographed, filmed, or recorded without their knowledge," and that a temporary ban would give time for "an informed debate" before permanent rules [1]. An expert group will propose those permanent national rules [1][2]. The government has not yet decided which devices would be covered; options include glasses with a camera and microphone and glasses with a camera and AI features [2]. AP reports that it is also considering exceptions so that "socially beneficial uses" can be permitted [3].

Others are moving in the same direction. Oslo schools and Equinor have their own bans, and Australia is considering one, according to [1]. AP notes that California's governor last week vetoed a bill that would have penalized recording people with smart glasses in private spaces [3]. The bill goes to the Storting "as soon as possible," the government says [3].

A Florida woman faces a felony charge after Anthropic reported messages she wrote to Claude to law enforcement [1][2]. According to the arrest report cited by [1] and [2], she wrote on September 26 that she would "shoot up" the Lee County Sheriff's Office, and a message the next day allegedly mentioned a new gun [2]. She has said she uses Claude like a "diary" [1].

The arrest report says Anthropic monitors chats for threatening content and, depending on severity, sends it for human review; in this case the review team reported it to police, according to [2]. Deputies detained her at home without incident [1][2]. Tom's Hardware says court records list a Sept. 30 charge under Florida Statute 836.10, which covers written or electronic threats of a mass shooting or act of terrorism, with arraignment set for Nov. 2 [2]. Anthropic has no comment on record [2].

Tom's Hardware quotes Anthropic's privacy policy, effective Sept. 10, as allowing disclosure to law enforcement where it believes it is "reasonably necessary to … prevent serious harm to any person or to property" [2]. It counts this as at least the third Claude conversation to reach police since August, after cases in San Antonio and San Francisco [2]. TechSpot links the case to pressure on OpenAI, which British Columbia has sued; TechSpot reports that OpenAI had flagged the shooter's conversations but never alerted police [1].

The practical point is simple. A chatbot conversation is not a private diary. Threats of violence can be read by a human reviewer and passed to police [1][2].