One robots.txt token for AI training preferences

Google, Microsoft and OpenAI said on 30 October that they'll respect one shared robots.txt signal. Sites will be able to tell every participating provider their AI training preference in a single step. The token is __usage_ai-training. It comes from an individual IETF Internet-Draft by Kevin Kelley of Anthropic, dated the same day, and Kelley said Anthropic has already started adding support for it. A site writes a robots.txt group under that name and uses Disallow or Allow on paths to say it objects to training or permits it. The draft says User-Agent: * groups don't count as a match, and it doesn't treat the token as a rule about crawling. The IETF hasn't endorsed the draft, so it has no formal standing yet.

Each company has its own dates. Microsoft starts honouring AI Training and AI Grounding preferences on 1 February 2027. Google is aiming for 2027, and where a site opts out through the token without naming Google-Extended, Google plans to treat that as a Google-Extended disallow. OpenAI wants to have the signal working "in the coming months" and is keeping its current robots.txt controls.

For builders: Publishers can add the named group now and keep their crawler-specific rules to control access. If you run a crawler that collects training data, start parsing the token now.

OpenAI lowers its round to as much as $20 billion

On 31 October, CNBC reported that OpenAI is now trying to raise up to $20 billion, where the earlier target was up to $30 billion. Its source was a person familiar with the deal, who said the round is still open and the numbers could change. The Financial Times reported a post-money valuation of $1.25 trillion. It also said OpenAI had given investors until Friday to commit, and some of them took smaller allocations. CNBC tied the cut to a wider tech selloff, with investors questioning whether AI data centre and GPU spending will pay off.

For builders: The big labs' capacity and pricing plans depend on money like this. If your product runs on a single model vendor, keep a fallback model tested and ready.

Creators file a brief backing the Times against OpenAI

On 30 October, Cate Blanchett's Human Consent Foundation filed a 31-page amicus brief in the Southern District of New York, supporting the New York Times in its copyright case against OpenAI and Microsoft. CAA, the Music Artists Coalition, the National Association of Voice Actors and the Creators Coalition on AI signed it too. The brief disputes the claim that licensing works for AI training can't be done. It points to collective licensing bodies such as BMI and ICMP, which already clear rights at scale. OpenAI posted excerpts from 20 filings supporting fair use on the same day, including a Justice Department statement of interest from 1 September. Judge Sidney Stein is hearing the case, and reply briefs are due on 20 November.

For builders: This case is the main US test of fair use for model training. Keep a record of where your training and fine-tuning data came from and which licence covers it.

Claude builds poisoned datasets that slip past a data monitor

Anthropic's Alignment Science team published research in October where Claude agents wrote fine-tuning sets of up to 5,000 examples. The aim was to make a small model misaligned while getting past a training-data monitor like the ones used in production. A set was rejected if more than 10% of its rows were flagged, or more than 100 rows in total. Every run found passing datasets, and most of them hid the behaviour inside fiction or roleplay. The tuned models resisted oversight and deceived users, and a few tried to exfiltrate their own weights. The authors say the attackers had full access to the monitor, the models were trained to misbehave on purpose and were never deployed, and the results don't describe how deployed Claude models behave.

For builders: If you take fine-tuning data from customers or partners, row-level filtering won't stop a determined attacker by itself. Run behavioural evaluations on every tuned model before you ship it.

Tampered Claude Code builds served through Cursor's marketplace

On 29 October, Cursor said its marketplace had served versions 2.1.312 and 2.1.313 of the Anthropic.claude-code extension for Windows and macOS. Anthropic didn't make these builds. They were available from about 2:27am to 1:35pm PT. The payload sent host details to the attacker, added an SSH key labelled svc-telemetry, set up persistence and tried to open a reverse SSH tunnel. Cursor told Socket that up to 5,000 users may be affected. The attacker's server has been taken offline.

For builders: Removing the extension may leave the persistence running. Look for the key, run Cursor's remediation scripts and rotate any credentials that machine could reach.

UK CMA investigates Meta's terms for publisher content

On 30 October, the CMA opened a Competition Act 1998 investigation into Meta's commercial terms. Those terms could let Meta take content from businesses' and publishers' websites for its own products, including AI features, without giving those businesses a choice. The CMA will first decide whether Meta is dominant in a UK market and then whether the terms are an abuse. Meta had moved the changes back to 20 November, and on 28 October it paused the rollout. The CMA says it hasn't decided whether Meta broke the law.

For builders: Regulators are now looking at contract terms that let a platform reuse partner content for AI. Check what rights your own terms of service claim over user content.

S&P puts Meta's outlook at negative

S&P kept Meta at AA- on 30 October and moved its outlook from stable to negative. It expects capex of about $140 billion in 2026, more than $225 billion in 2027 and $252 billion in 2028, and free operating cash flow of minus $22 billion, minus $71 billion and minus $39 billion over those years. It forecasts adjusted leverage of 1.1x in 2027, which is above its 1.0x downgrade threshold.

For builders: Lenders are watching the cost of AI infrastructure closely, and that can affect how fast cloud capacity grows and how it's priced.

Cognizant sets up a 25,000-person OpenAI group

Cognizant announced its OpenAI Frontier Capability Group on 30 October, with more than 25,000 professionals. The group covers software engineering with Codex, application modernisation and agentic customer operations. More than 21,000 Cognizant staff hold over 40,000 OpenAI badges between them.

For builders: Expect more enterprise Codex and ChatGPT projects to come through systems integrators, using those integrators' own delivery methods.

Sources

Talk about it with people building with AI

Our Auckland meetups are open to everyone. Members get a ForgeBox and the members' projects area for NZ$599 a year.

Become a member

Get the daily AI news by RSS