Lead story / research

Two AI tools link facial photographs to cancer survival risk

Two AI tools reading facial photographs captured different signals associated with cancer patients’ survival, Mass General Brigham researchers report. Combining their biological-age and direct survival-risk estimates improved survival-model fit in an external patient group. Published October 19 in The Lancet Digital Health, the findings remain research—not a clinical tool.

Both tools build on FAHR-Face, trained on more than 40 million facial images. One estimates biological age; the other estimates survival risk directly. Their predictions were only weakly correlated, and each remained independently associated with survival when analyzed together, suggesting they captured complementary information rather than the same signal twice.

In a separate test group of more than 38,000 cancer patients, those assigned a biological age above their chronological age had higher mortality risk. Scientific American also reported substantial mortality within 90 days in the highest-risk subgroup of a 3,820-patient test group. That subgroup finding is not a measure of overall model accuracy.

Clinical use is a separate question. The survival model’s training population was majority white and drawn from the Boston area, and researchers say broader data and clinical testing are needed. Radiation oncologist Raymond Mak wants the tools to support, not replace, physician judgment. Bioethicist Matthew Liao warned that a mistaken estimate could lead to needed care being withheld, and raised possible misuse by insurers or employers.

security risk

Hacker used AI agents to break into sites at over 100 companies, Reuters says

Reuters reports that a hacker used AI agents to probe more than 1,200 sites and breach websites at over 100 companies. A researcher watched the agents through the attacker’s unsecured Cairn interfaces, where they coordinated work and pursued intrusions for hours. Reuters could not establish the extent of data theft. NorForge’s roughly $5,000 estimate covers AI tokens, not a verified operating budget, and carries considerable uncertainty.

startups

Apate says its AI decoys held 2.5 million conversations with scammers

Apate’s agents pose as potential victims to extract bank accounts, crypto wallets and other scam intelligence for banks and telecoms. TPG Telecom diverted nearly 590,000 calls to the service, according to an Apate co-founder. The company also cites a 76% decline in publicly reported scam losses at Commonwealth Bank. Experts say independent analysis is needed to separate Apate’s effect from other safeguards and scams shifting to new targets.

research

MIT says Claude agents improved at tasks after exploring a virtual island

In a study that has not been peer-reviewed, 13 Claude Opus 4.7 agents spent 30 hours on self-chosen island activities. They later outperformed fresh copies on tasks including moving objects to a summit. Removing information from their persistent text notes weakened performance, linking retained discoveries to later success. The results suggest exploration can build reusable skills, but do not establish genuine play; prompting or imitation may explain some behavior.

Also worth knowing

•  KG Growth Partners says a client is weighing a data center in San Francisco’s closed mall

•  Army is developing AI agents for cyberattacks with little oversight, The Intercept says

•  Judge keeps former OpenAI engineer detained, citing flight risk

•  FDA clears an AI tool to assess breast cancer recurrence risk before surgery

•  OpenAI releases a public coding test for 20 programming languages

•  OpenCode Console breach exposes AI keys and triggers unauthorized charges

•  AI agents finish under half of healthcare tasks in Stanford simulations

•  fal launches an AI video model with a claimed 22-fold speed advantage

•  Cornell wins $600,000 to make AI forecasts of human choices more explainable

•  Florida sheriff says scammers cloned a captain’s voice for impersonation calls

Daily tool drop

6 AI tools worth knowing today

The Internet Had a Point

The Goalposts Keep Moving on LLM Math

Reader check-in

Help shape tomorrow's briefing

One click tells us what to keep, improve, or tighten.