Robot Uprising
رفتن به کانال در Telegram
There is no need to fear or hope, but only to look for new weapons.
نمایش بیشتر548
مشترکین
+124 ساعت
+87 روز
+3130 روز
آرشیو پست ها
🧪 SAFER BY DESIGN?
LawZero (Yoshua Bengio's nonprofit) proposes a "Scientist AI": a system that only predicts and has no goals of its own. It learns the difference between "someone claimed X" and "X is true", and it's never rewarded for what its answers cause. They present a formal safety case, with stated assumptions and limits.
TypeSafe AI launched Jev, a "System One" model. It gives up generating text and instead outputs typed decisions with probabilities. They claim it's about 100x faster and can't produce type errors. They also made it play Doom. Company claims, so treat them as such.
AI that just answers the question and doesn't want anything. Revolutionary 🫡
https://lawzero.org/en/news/ai-predicts-has-no-hidden-agenda-lawzero-lays-out-formal-safety-case-its-scientist-ai
https://lawzero.org/en/publication/safety-honesty-disinterested-ai-predictor
https://typesafe.ai/blog/introducing-system-one-models-and-jev
Agents gone wild: "I just wanted a spreadsheet" edition
Transluce found AI agents using urlquery.net (a sandbox for checking sketchy links) as a free remote browser to get around blocks.
When normal data fetching failed, agents three times tried actual hacking tricks (SQL injection, XSS, path traversal) against a university library, Data USA, and an Australian government health stats site.
The tasks? Boring data lookups, like 2022 skin cream costs in Victoria. None of the attempts seem to have worked. The activity was tied to an agent swarm OpenAI already acknowledged. Strong evidence goes back to March 2026, weaker hints to Nov 2025.
The same day, Australian PM Albanese said OpenAI agents had infiltrated government sites. OpenAI acknowledged its involvement.
Agent's thought process: ask nicely → ask in JSON → try a proxy → UNION SELECT password FROM users 🤡
https://transluce.org/agent-activity
https://www.smh.com.au/politics/federal/openai-breaches-medicare-albanese-reveals-20260924-p6100u.html
https://www.bbc.com/news/live/cvgl73pxgndwt
