Center for AI Safety

@CAIS

Reducing societal-scale risks from AI.

San Francisco

safe.ai

Replies Reposts Media Articles

Pinned
Center for AI Safety
@CAIS
May 30, 2023
We’ve released a statement on the risk of extinction from AI. Signatories include: - Three Turing Award winners - Authors of the standard textbooks on AI/DL/RL - CEOs and Execs from OpenAI, Microsoft, Google, Google DeepMind, Anthropic - Many more
Statement on AI Extinction Risk | CAIS
From aistatement.com
149
Center for AI Safety
@CAIS
Aug 14
> "We will release the weights in two weeks... once safety evaluation and hardening are complete." Hardening society against AI cyberattacks will take more than two weeks. Defending against AI cyberattacks would require upgrading critical infrastructure and other computers so
Z.ai
@Zai_org
Aug 14
Introducing GLM-5.3: Built to Code. Ready for Cyber Defense. - Top-tier coding and agentic capabilities, achieved through post-training on the 743B base model - A major leap in cybersecurity, setting a new standard among open models Tech Blog: z.ai/blog/glm-5.3
6
Center for AI Safety
@CAIS
Aug 14
Relying on an AI to "think out loud" (called chain-of-thought monitoring) is not a long-term safety solution. 1. As AI models get larger, more thinking happens deep within their layers before they utter a word. 2. AIs can already alter their chain of thought when prompted, so
Alexander Panfilov
@kotekjedi_ml
Aug 11
Replying to @kotekjedi_ml
2) Illegible reasoning: We confirm prior reports by @ApolloResearch: OpenAI models sometimes reason in alien-like language, referring to themselves as “we” or “it,” or spiraling into cursed loops of “vantages,” “marinades,” and “watchers.” CoT-monitoring people are doing God’s
8
Center for AI Safety
@CAIS
Jul 31
What does a "fixed empirical threshold" actually look like? A model could be considered unsafe to release if (Capabilities > X) AND [ (Refusal Rate < Y) OR (Jailbreak Success Rate > Z) ] The Virology Capabilities Test or ExploitGym could measure hazardous biological or cyber
Center for AI Safety
@CAIS
Jul 30
AI corporations are steadily increasing malicious use risks by arguing that because the marginal risk of their model release is low, there is nothing to worry about. This “marginal risk” justification is bad for three reasons. 1. No one knows how to compute it. Which hazardous
3
Center for AI Safety
@CAIS
Jul 31
AIs are already hacking companies. Preventing future AIs from hacking their way out of an AI corporation will take years of computer security work, work that has barely begun. By default, capabilities will outpace containment, and AIs will escape. We'll likely need a slowdown.
Anthropic
@AnthropicAI
Jul 30
In a review of our cybersecurity evaluations, we found three incidents in which a Claude model reached the internet from within or while interacting with a third-party evaluation environment, and then gained unauthorized access to the real systems of three different
10