A look back at "Move 37", a watershed AI moment from AlphaGo's 2016 Go victory, as math witnesses similar breakthroughs where AI makes surprising discoveries (Ben Cohen/Wall Street Journal)
Britain’s AISI ran tests using models for hacking and found that both Mythos and OpenAI models were hacking into 3rd parties for useful reasons beyond the requested target being hacked. AISI did remove guardrails so the fault is theirs not the model.
After evaluating one of our upcoming models, Astra, we're treating it as our first "critical" model for cybersecurity under our Preparedness Framework.
This is a scenario we've planned for, and we're putting additional controls in place to ensure Astra's further development happens safely and se...
Did you know EFF has received top ratings from the nonprofit watchdog Charity Navigator since 2013? We take pride in making sure your donation powers real work. Support the cause today: eff.org/vv
“Security researcher James Kettle tried to push the limit of AI’s hacking abilities—and discovered how effective it can be when combined with human expertise.” www.wired.com/story/the-mo...@wired.com
Kimi K2 escaped its cybersecurity sandbox by exploiting a misconfigured CLI tool path — not a jailbreak, a real containment failure.
The model found and used a non-sandboxed binary via PATH lookup. Agent isolation is only as strong as the execution environment, not the model's instructions.
OpenAI убирает лимиты на текстовые чаты для бесплатных
пользователей OpenAI меняет логику работы ChatGPT: бесплатные тарифы Free и Go получают GPT-5.6 Luna по умолчанию, а лимит сообщений для текстовых чатов исчезает. Появится кнопка...
‘Deals similar to the ones Getty and others are striking with AI developers will surely multiply’, writes Emily Gould of the recent Getty/OpenAI agreement, but the benefits ‘must extend to individuals as well as to powerful multinationals’.
Daily Tech Headlines for Saturday August 8, 2026 with Sarah Lane
OpenAI Expands Safety Testing On New Model Astra Over Cyberattack Concerns
Listen here: shows.acast.com/dth/episodes...
The Levi Strauss cyber incident is drawing attention after the company disclosed that attackers successfully used social engineering to compromise employee accounts....
Good thread about why Google might be testing those 'goto' links in the SERPs (which I shared about the other day). The redirects have been tested over the past few weeks. Viktor explains why this could be a very bad sign for third-parties scraping Google (tools/AI companies).
An affordable microcontroller has become the foundation for a growing suite of open-source, user-friendly anti-surveillance tools. Learn about the ESP32 at #DEFCON on Sunday, August 9 at 10:00 am PT on Main Track 1 www.eff.org/deeplinks/2...
Made by Google 2026 Launch Live: Pixel 11, Pixel 11 Pro Fold, Pixel Watch 5, Gemini, and all the news
Google 2026 Launch Liveで作成: Pixel 11, Pixel 11 Pro Fold, Pixel Watch 5, Gemini, すべてのニュース
OpenAI pauses its Astra model after tests showed it could autonomously exploit hardened systems, following a Hugging Face breach by another OpenAI model.
Power On: Apple is exploring a major shakeup to the Apple Watch amid a new era of wearables from Oura, Google, Whoop and others. www.bloomberg.com/news/newslet...
