Research done as part of the Singapore AI Safety Fellowship shows how Anthropic’s AI model can embed hidden messages into images using Morse code, object rotations, colors, and heights. The researchers also found that separate sessions of Fable 5.1 could decipher about 78 percent of the hidden messages based on a screenshot.
Charlie Warzel at The Atlantic interviewed Ondřej Surý, the man almost single-handedly responsible for packaging up the patch for this week’s vulnerabilities. It’s a fascinating conversation, from Surý‘s odd nonchalance about saving the web to his concern that for-profit companies are exploiting open-source. But he hopes that this AI-assisted discovery will lead to a change:
… Everything’s moving fast now, and it’s taking a toll on the people who are defending the stuff. At the same time, maybe this is the turning point where we actually start caring about this … I might be wrong, but I’m an optimist that people will finally realize that we can’t go like this forever. This has been a disaster waiting to happen for a really, really long time.
Tomek Korbak, Mikita Balesni, and Jasmine Wang tell The Wall Street Journal that OpenAI terminated them after finding disappearing messages on Signal in a chat with Ryan Greenblatt, the former chief scientist at AI safety evaluator Redwood Research. OpenAI reportedly accused Greenblatt of leaking details about its Astra AI model to The Information, which Greenblatt has denied.
OpenAI tells the WSJ that the researchers were fired for violating “clear policies for handling sensitive information,” and “not for raising safety concerns or speaking out.”
[The Wall Street Journal]
A rise in AI-powered decomps, recomps, and ports has thrust the community into chaos.
And a red pill meme. You can see the post on X here.
Epic CEO Tim Sweeney said last week that the company is undergoing a “huge pivot” to “retool for maximum AI acceleration,” and it announced a “vibe coded” update to its PC launcher last week in preview that had some issues on the first day.
From July 21st to August 11th, two employees from METR and Redwood Research conducted a “pilot” to stress test the system, and METR published a high-level overview of their findings, as reported by CNBC. One exploit would have allowed agents to “prevent hundreds of recent actions (including the actions involved in the exploit itself) from being monitored.”
OpenAI “has stated” that the vulnerabilities identified “that impact their production safeguards stack” have been patched, METR says.
Yes, of course there are pay package incentives in play for Musk. But that’s not all.
