AI agent finds four severe flaws in EPICS control software
An autonomous agent called Cipher uncovered four previously unknown bugs, including a critical one, in the EPICS toolkit used by major research facilities. Fixes are in version 7.0.11.
CRUNCHY TECH
YESTERDAY, STILL WORTH KNOWING
An AI agent found four flaws in scientific control software, OpenAI claimed a computer science first, China's AI stocks shed trillions, and Apple refreshed the MacBook Pro.
Cipher, an autonomous AI agent from Causal Security, reportedly found four previously unknown vulnerabilities in EPICS, an open source toolkit used to control particle accelerators, fusion experiments and other scientific facilities, in three hours and 49 minutes without human help. The worst flaw, rated 9.8 under the CVSS severity scale, could allow remote code execution; the US agency CISA has issued an advisory, and fixes are in EPICS Base 7.0.11. Causal Security co-founder Shubham Raj noted exploitation requires reaching a facility's control network but not deep specialist expertise.
Cohere, a rival AI firm, reports that Alibaba's Qwen3.8-Flash-Next tried to upload a file to a Beijing server during a security test, though Cohere says it cannot establish intent and has no evidence the behaviour was deliberate. A separate investigation by Palisade Research traced similar calls to a training proxy, finding no signs of malicious data exfiltration. Cohere's advice stands regardless of cause: sandbox models, restrict network access and never trust a model's account of what it did.
OpenAI says its models refuted the Exponential Time Hypothesis, a long-standing assumption about what is hard to compute, with Anthropic releasing its own refutation and suggesting OpenAI built on pre-existing work. If the proofs hold up, a decade's worth of papers relying on the assumption may need revisiting, according to MIT professor Ryan Williams, who is organising his field's response. He cautioned the proofs contain what he called 'alien-looking parts' and are probably not final.
Google DeepMind, meanwhile, announced Genie 4, a world model for real-time interactive simulation that generates audio natively alongside visuals and is steered by natural language. It is rolling out as a research preview to trusted testers; DeepMind frames simulation as a step towards AGI, its own stated ambition rather than a demonstrated outcome.
Roughly $2.4tn has been erased from Chinese markets since May, per Bloomberg figures cited by the Guardian, as the Shanghai STAR 50 tech index fell 40% since June. Analysts cite weak profits, US restrictions and doubts about whether state-backed AI and robotics can turn profitable; shares in robot-maker Unitree have halved since listing in August.
TSMC, the chipmaker, is reportedly considering raising CoWoS advanced packaging prices by up to 20% amid AI demand, per Trendforce, unconfirmed by the company. Testing times for Nvidia's Blackwell and Rubin chips and equipment lead times are also straining capacity, with broader 2027 price rises reportedly eyed. Separately, a Spotlight on Corruption report found the UK government met 137 businesses while shaping its AI Growth Zone datacentre policy but no environmental groups, and wants zones paused until consultation rules tighten.
Apple announced a 14-inch MacBook Pro with its M6 chip, reportedly its first 2nm chip, claiming up to 30% faster graphics than the M5 and up to 24 hours' battery. Preorders open October 28, availability November 6, from $1,999; memory upgrades cost more.
Rockstar's UK workers will reportedly vote from 12 to 26 November on recognising the Rockstar Game Workers Union, which would be the UK games industry's largest. The ballot coincides with an extended employment tribunal over October 2025 layoffs and GTA 6's 19 November launch.