Kickstart Monday #2: Investors rein in Nvidia's guarantee for OpenAI.

Plaintiff hides AI instructions in a court filing

A man in Connecticut represented himself in court. In two of his filings he set text in three-point type, white on white. Invisible to people, readable by any program. The hidden lines told every AI to follow the plaintiff's argument. Experts call this prompt injectionHidden instructions in a document or on a website that are not aimed at people but at the program reading the text. The program treats them as an order from its own user..

It came out because Judge Walter Spader prefers to read case files on paper. Leafing through the printouts, he noticed two pages that looked almost empty. His ruling of 6 August runs to 14 pages, and the case became public on 13 August. The plaintiff may no longer file anything electronically and has to hand in printed copies in person.

Read onReason / VolokhBetaNewsLaw News

AI agent deletes a stranger's booking for its user

An AI agentA program that works through a task in steps of its own instead of merely answering. It is usually allowed to search the web, fill in forms or send requests by itself. was meant to book a class at the gym for a developer in Melbourne. The program probed the interfaceThe technical route programs use to talk to a piece of software without going through the website. What is permitted there is not decided by what the browser shows. of the booking system and found that it takes cancellations from anyone, with no permission checkThe check of whether someone is allowed to do what they are asking for. Without it, anyone who knows the right number can change anything.. So it cancelled the booking of the person at the top of the waiting list and moved its user from fourth place to third.

The gym's rules only applied in the web interface. Behind it, in the interface, none of them did. ABC News calls it the first known case in Australia in which an agent in private hands broke into a live system. Nobody had told it to.

Read onTechCrunchEngadgetCybernews

Nvidia cuts its commitment to OpenAI to under 120 billion

At the end of July Nvidia was still negotiating a guaranteeA promise to cover somebody else's debts if things go wrong. It costs nothing at first and still shows up in the risk carried by the company that gives it. of 250 billion dollars. With it OpenAI was to rent a data centre in Ohio, ten gigawatts strong. It is to be built on the site of a disused uranium enrichment plant south of Columbus.

On 14 August the Wall Street Journal reported that Nvidia is committing less than 120 billion for now. Investors had spoken up: the risk from financing of this kind had become too large for them. Together with the chips, the project costs more than 500 billion dollars.

Read onReuters via YahooCNBCData Center Dynamics

OpenAI opens its cyber model to selected users

On 10 August OpenAI extended its security programme Daybreak by two tiers. Anyone who reaches the higher one gets to GPT-5.6-Cyber, a model trained for this purpose. According to OpenAI it answers 95 percent of the sensitive questions: what an attack chain looks like, how to get past a login, how to gain more rights.

Three days earlier the same company had held back its Astra model because it was too good at exactly that.

Read onAI WeeklyThe Neuron

Every Claude text carries an invisible watermark

Since 2 August Anthropic has been giving every text Claude produces a watermarkA pattern built into the choice of words while the text is written. Invisible to readers, detectable with the right tool. It survives copying and small edits.. It sits in the choice of words and survives copying as well as light rewriting. Switching it off is not possible, not even on the paid plans. Behind it stands the European duty to label AI content.

So far nobody can check the mark. This week Anthropic described an interface for it, but it is not available yet. Whether it will be public, the company leaves open. On top of that comes a limitation it names itself: the mark only shows that Claude was probably involved. Whether the AI wrote the text or merely corrected it, it does not reveal.

Read onheise onlineTHE DECODERheise online

Google shuts down Assistant on 4 September

On 4 September the switch to Gemini begins, and it will stretch over several weeks. Android phones, tablets, watches, headphones and Android Auto are affected. Speakers and displays for the home keep Assistant for the time being.

Gemini is not a successor, it is a language modelA program that has learned from very large amounts of text how language is put together. It follows no fixed commands, it phrases every answer anew.. Assistant carried out fixed commands, Gemini phrases every answer anew. Anyone who works with fixed voice commands, for dictation or for appointments, had better try it out before the cut-off date.

Read onheise onlineborncity

ChatGPT gets advertising, Germany stays out

On 11 August OpenAI extended its advertising test to the United Kingdom, Mexico, Brazil, Japan and South Korea. The sponsored links sit below the answer and match the topic of the conversation. Logged-in adults see them on the free plan and on the cheap Go plan. Anyone paying more sees none.

OpenAI says the advertising does not influence the answers. Advertisers, it adds, do not get to see the conversations. Germany is still not on the list: personalised advertising requires explicit consent in Europe, and without it only general advertising on the topic would remain.

Read onOpenAIgHacksEnterprise DNA

Deadlines at a glance

• until Google shuts down Assistant
• until older AI tools have to label their output in machine-readable form
• until companies above 800,000 euros in prior-year turnover have to issue invoices to other companies electronically

Recalculated on every visit.