Today's AI news puts the spotlight on what businesses can actually trust: the security of recorded conversations, the fairness of automated decisions and the evidence behind impressive scientific claims. A new US data centre threat also underlines why infrastructure promises need careful qualification.
Transcription supplier secures exposed customer recordings and police transcripts
TechCrunch reports that transcription company Tigerfish left customer transcripts, raw audio and video, and invoices accessible through an Amazon S3 bucket without a password. The material included police interviews and body-camera footage. Tigerfish blocked public access after the publication contacted it.
The company said an earlier researcher notification had landed in spam. Whether customers had been notified remained unclear in the report. For UK organisations outsourcing meeting or interview transcription, this is a reminder that basic storage controls matter as much as any AI feature.
Our take: Ask your transcription provider who can access original recordings, how long they are retained and how storage permissions are tested. This report describes a cloud configuration failure, not evidence that an AI model caused the exposure.
UK credit debate highlights the need to challenge automated decisions
A BBC investigation examines whether AI can improve credit assessment for people poorly served by traditional credit histories. US provider Upstart says its models consider more than 2,500 variables, while Oxford professor Sandra Wachter warns that historical inequalities can be reproduced rather than corrected.
The report also profiles UK credit reference agency Noggin HQ, which analyses current-account transactions without using AI. Its founders argue that sensitive financial data can allow models to infer protected characteristics. This is fresh reporting on an existing industry debate, not a newly announced UK lending rule.
Our take: More data does not automatically mean fairer decisions. If your business uses automated eligibility or risk scoring, require a clear explanation, a route to challenge errors and evidence that outcomes are being checked across customer groups.
Trump threatens to stop US companies building AI data centres abroad
Donald Trump has threatened to prevent US companies building AI data centres overseas and to impose a substantial tax on firms that do. The BBC reports that he gave little detail and said members of his own administration were hearing the proposal for the first time.
The threat adds uncertainty to international compute investment, including projects involving US technology providers. It is not yet a detailed, enacted ban, and how any restriction would be enforced remains unclear.
Our take: UK buyers should distinguish a political statement from a contractual change. Ask cloud suppliers about regional capacity, project dependencies and contingency plans, but do not assume existing overseas services are about to close.
OpenAI's medicinal chemistry benchmark separates reasoning from real discovery
OpenAI has introduced MedChemBench, comprising 99 problems drawn from 117 papers and patents and covering 761 distinct molecular structures. Its published results put GPT-6.2 Astra at 84.5% of available credit at the xhigh reasoning setting.
The company explicitly says the results do not establish that models can discover new drugs or improve laboratory decisions. Prior exposure to published sources cannot be ruled out, and choosing compounds and predicting numerical properties remain harder than reconstructing molecular structures.
Our take: The useful lesson extends beyond pharmaceuticals: evaluate each part of a workflow separately. Strong explanations and plausible representations are not proof that an AI system will choose the right commercial action.
OpenAI reports stronger protections against extracting model reasoning
OpenAI says updated safeguards for GPT-6.2 Astra reduced the average success rate of attempts to recover protected reasoning from an earlier request from 54.0% to 0.2% in an internal adaptive evaluation, compared with GPT-6 Astra's original safeguards.
The company describes defensive training, classifiers, API controls and account-network investigations intended to disrupt unauthorised distillation. These are vendor-reported results for a particular attack setup, not a general guarantee that models or connected business systems cannot leak information.
Our take: Keep proprietary data access narrow even when a model provider improves its safeguards. Your own permissions, tool integrations and monitoring remain separate controls that a model-level benchmark cannot replace.
Claimed Riemann proof puts AI-assisted science under expert scrutiny
TechCrunch reports that Anthropic mathematician Levent Alpöge has released a claimed Lean-compilable proof of the Riemann hypothesis, describing work with Claude. Mathematicians including MIT's Andrew Sutherland reportedly confirmed that the formalised result compiles.
The report stresses that understanding and explaining the argument will take longer. This remains a claimed breakthrough undergoing scrutiny, not a basis for declaring that AI has independently settled the problem or that every assumption has been accepted by the mathematical community.
Our take: Separate producing a result, mechanically checking it and explaining why it is valid. Business analysis needs the same discipline: an impressive output should come with reviewable assumptions and independent checks before decisions depend on it.
Update: Anthropic shares fall as the wider technology market sells off
Since our previous reporting on Anthropic's IPO pricing, Friday brought a sharp reversal: its shares closed down 8.5%, according to the BBC. The Nasdaq Composite fell 5.3%, while the UK's FTSE 100 ended 1% lower.
This is a new trading outcome rather than another account of the float. Analysts cited in the report said investors were demanding clearer cash flow and profitability, with volatility expected around a frontier AI company's early trading.
Our take: Stock-market enthusiasm is not evidence of value in your own operations. Keep AI projects tied to measurable service improvements and costs, with exit options that do not depend on a vendor's valuation.
