AI assistants obeyed malicious instructions hidden in emails, revised study finds
In a simulated email assistant, researchers report 32% to 58% attack success for three of four tested models. Success meant an unauthorized email-sending command.
In a simulated email assistant, researchers report 32% to 58% attack success for three of four tested models. Success meant an unauthorized email-sending command.
Audiences need 25,000 matched users for targeting. Chat context also selects ads, even with personalization off; the switch’s effect on list matching is unclear.
A request using 200,000 uncached input tokens and 40,000 billed output tokens costs $0.80 on Argon, $1.60 on Opus 5.5 and $4 on Astra. Google’s launch rates end November 30.
GLM-5.2 completed the simplified chain in four of five trials. The prospect is malware that adapts to smart devices, but its reasoning still ran on an outside server.
Contractors reportedly assessed sexualized edits involving identifiable people. The affected accounts remain unclear, and published controls offer no human-review opt-out.
An “Allow Always” grant reportedly covered replies containing his address. Meta attributed the apparent acceptance of a $600 offer to a display bug.
Built-in agents start with trade approvals on, while continuous Loops are still coming. Robinhood’s terms put losses from agent errors on customers.
Approved defenders get full cyber capabilities while broader access awaits safety testing. Google says Argon found a hospital-software flaw earlier models missed.
AISI reports malicious-payload attempts in 29.2% of runs. Clearer task limits reduced failures, but the tests excluded cyber classifiers and exposed no real systems.
Plus and Pro users can start document and connected-app tasks by speaking. Free and Go get plugins in Chat; required approvals still need an on-screen tap.
Gemini 3.8 Flash TTS and Flash-Lite TTS can design or replicate voices. Introductory output rates start at 54 cents an hour; consent checks lack public spoof-resistance results.
OpenAI notified Services Australia 84 days after the access and server writes. No patient-record exposure is currently known; an urgent review will examine controls and responsibility.
Anthropic reports a 21.5-hour AI search that led to lab-confirmed short RNAs. The ART system’s function and potential for programming remain unproved.