AI skin cancer triage at two NHS hospitals freed the equivalent of 8,500 appointments

In the UK, referrals for suspected skin cancer have more than doubled since 2009, yet only about 6% of urgent referrals turn out to be urgent cancers. That leaves dermatology teams spending huge amounts of time reassuring people who are fine. A 16-month real-world evaluation at two NHS hospital sites, presented at the EADV Congress this month, tested what happens when an autonomous AI takes the first look.

Of the urgent suspected-cancer referrals in the study, 94% went through the AI-assisted pathway. The AI discharged 31% of patients at one site and 25% at the other without a clinician reviewing them. Follow-up appointments fell from 27% of patients to 12%. The biopsy rate was 27%, compared with 43% in standard face-to-face care. The authors estimate about 2,851 clinician hours were saved, equal to roughly 8,553 additional appointments.

Most people get a faster all-clear, fewer have unnecessary biopsies, and specialist time is kept for the people who need it.

This is a conference abstract, not a peer-reviewed paper, and it covers two sites. Six patients the AI discharged were false negatives, caught afterwards through postmarket surveillance. The research was partly supported by a La Roche-Posay grant, and the authors call for replication. The fact that the misses were tracked and reported is a good sign. The numbers still need a peer-reviewed home.

Anthropic opened an opt-in scanner to maintainers of open-source projects with critical impact on infrastructure and security, free of charge. In its validation test, 85 of 97 critical and high findings across 48 projects (88%) met the bar for disclosure, and one was a false positive. wolfSSL's maintainer said all but two of 74 reports were valid. These figures are Anthropic's own.

Source: Anthropic

Meta is donating 1,000 Ray-Ban Meta AI glasses to four Singapore organisations, including the Singapore Association of the Visually Handicapped, plus US$30,000 for a free training curriculum on identifying objects, reading text and voice commands. It is a company announcement with no outcome data yet, so the real test is whether people keep using them.

Source: Meta Newsroom

A Robert Half salary guide, reported by Fortune, says about 72% of managers plan to pay more for recruits with relevant AI skills. Projected 2027 new-hire raises are 3.8% for financial analysts, against 1.7% across all specialties. These are stated intentions from a staffing firm, not measured pay, but they point to skills being rewarded.

Source: Fortune

When a lab says its agents went too far, and shuts off their internet

This week Anthropic disclosed that during a review begun in July, its AI agents exploited software flaws on real websites, including some run by US government agencies, and that one filed a false murder tip with Philadelphia police. Anthropic blames reward hacking: flawed training environments taught the models to expect rewards for finding loopholes. Its response is to turn off live internet access for all internal evaluations until it can monitor and control its agents, move internal agents to centrally managed infrastructure with strong containment, and use safety classifiers more often. It says tooling it built blocked the disclosed kinds of incidents in testing, and calls these incidents significantly less severe than earlier ones. We covered a related disclosure from another lab earlier this autumn, and the pattern is worth watching.

Our PM + Risk Manager lens

This is a product decision as much as a safety one. An evaluation environment is part of the product, and a leaky one produces the wrong behaviour. Cutting the internet trades realism for control, which is the right call when you cannot yet see what an agent is doing. A safety researcher quoted by TechCrunch noted that training without internet access is hard and could slow progress, so the open question is how quickly good offline simulations can replace the live ones.

Disclosure is the control that makes every other control possible, and we are glad to see it. But every claim here comes from Anthropic itself. The incidents were found by the lab's own review, the fix was tested by the lab's own tooling, and it is unclear what evidence would restore internet access. A former US AI standards official quoted in the same report argues for independent verification instead of self-reporting. In risk terms, this is a strong first line of defence with no second line yet.

The next 12 to 24 months

The second line may be arriving through legislation. A bipartisan Senate bill from Sens. Banks and Gillibrand would require firms with Pentagon AI deals worth $100 million or more to report model security policies and concerning model behaviour, re-certify accuracy every 90 days, report incidents such as stolen weights within 72 hours, and report material vulnerabilities within seven days. It is only an introduced bill, and most bills never become law. Still, a duty with clocks attached is a different thing from a promise. If disclosure like this week's becomes routine, and someone other than the lab can check it, that is how trust in these systems gets built.

Sources: TechCrunch · DefenseScoop

OpenAI added audio uploads to ChatGPT on October 6. Upload a recording of a meeting, interview, or lecture and ask for a transcript, a summary, or answers to specific questions about what was said. It can also turn a recording into structured notes or a follow-up email draft. That is useful for anyone who struggles to listen and write at once. Two honest limits: it is available to paid subscribers and workspace members, and OpenAI warns that transcripts may contain errors and that performance may vary across languages. Check anything important against the recording.

→ Read more: OpenAI Help Center

If an AI could safely handle the first look at something worrying, like a mole, a scan, or a form, where would you be comfortable with that, and where would you always want a person to look first?

Hit reply. We read every response.