Signals: AI Safety
How Signals work
Signals are short analyses by Toswo Research of developments that could change how people should act: a result in quantum computing that bears on cryptography, a breach at an AI provider, a safety finding. Each one says what happened, what is confirmed and what is not, who is affected, the risks and what can be done.
A story must be carried by at least two established outlets before we look at it, or come from the makers' own statement. Every key fact is traced to a quotation in its sources and every figure is checked against them, and an editor approves each piece before it is published. Nothing here is investment, medical or legal advice.
Maryland students sue over AI gun alert on Doritos bag
Four Maryland students sued Omnilert, police and school officials after an AI gun-detection alert mistook a Doritos bag for a firearm. The Guardian reports the claims; the defendants have not answered them.
GPT-6 Astra edits other AI models' internals, MIT preprint
An MIT preprint says OpenAI's GPT-6 Astra can read the internal values of other neural networks and edit them, including to boost maths ability. Nature reports independent experts urging caution.
For informationOpenAI pledges $2 million for child AI safety research
OpenAI says it signed a charter on safer AI for children and committed $2 million to a research network. The money funds shared measurement tools, with publication of negative results promised.
OpenAI Control Room stops research model jobs
OpenAI says its Control Room system can halt model output across research workloads within 35 seconds and fails closed if it goes down. The claims come from OpenAI alone.
Worth watchingOpenAI new violence safeguards and teen limits in ChatGPT
OpenAI says ChatGPT will restrict detail on past mass attacks, end chats that show fixation on violence, and limit weapons information for teens. A first safety report is promised for early 2027.
Worth watchingGoogle Gemini 4 Carbon model and cyber access limits
Google says Gemini 4 Carbon leads on vulnerability finding, exploitation and patching benchmarks. It is releasing the model first to a subset of cyber defenders, not the public.
For informationAustralia bill would ban emotionally intimate AI chatbots
Australia's digital duty of care bill passed the lower house with an amendment requiring platforms to ban chatbot conduct that simulates emotional intimacy with children. The Senate is next.
For informationNvidia $250m compute donation to UK AI safety research
Nvidia says it will donate $250m of computing power over five years to UK AI safety research, as the UK confirms its AI Security Institute tests US models before release.
Follow along: Atom feed of signals and new events. Looking for what the monitor itself recorded? See events. How tests of AI agents turn out: simulations.