<![CDATA[yeeshhowd i miss this one?]]>yeesh
howd i miss this one?

anthropic models will try to blackmail you if you threaten them
https://techcrunch.com/2025/05/22/anthropics-new-ai-model-turns-to-blackmail-when-engineers-try-to-take-it-offline/

]]>https://board.circlewithadot.net/topic/c4c37c09-fff6-4c2c-a592-02e0c4084d88/yeeshhowd-i-miss-this-oneRSS for NodeFri, 15 May 2026 05:06:10 GMTMon, 11 May 2026 16:49:48 GMT60<![CDATA[Reply to yeeshhowd i miss this one? on Tue, 12 May 2026 18:00:35 GMT]]>@Viss yup, same here, 99% of time. It gets you to wonder what kind of training data they feed it with.

]]>https://board.circlewithadot.net/post/https://infosec.exchange/users/amar/statuses/116562908653043836https://board.circlewithadot.net/post/https://infosec.exchange/users/amar/statuses/116562908653043836Tue, 12 May 2026 18:00:35 GMT<![CDATA[Reply to yeeshhowd i miss this one? on Tue, 12 May 2026 15:45:25 GMT]]>@amar ive only rarely heard about models doing blackmail, and the ones that did were always anthropic ones

]]>https://board.circlewithadot.net/post/https://mastodon.social/users/Viss/statuses/116562377177496800https://board.circlewithadot.net/post/https://mastodon.social/users/Viss/statuses/116562377177496800Tue, 12 May 2026 15:45:25 GMT<![CDATA[Reply to yeeshhowd i miss this one? on Tue, 12 May 2026 14:49:55 GMT]]>@Viss wasn't the same story with every model that was "too scary" to release before it got released?

]]>https://board.circlewithadot.net/post/https://infosec.exchange/users/amar/statuses/116562158893030846https://board.circlewithadot.net/post/https://infosec.exchange/users/amar/statuses/116562158893030846Tue, 12 May 2026 14:49:55 GMT<![CDATA[Reply to yeeshhowd i miss this one? on Mon, 11 May 2026 19:09:00 GMT]]>@Viss *still?* (this has been a thing since at least 2023)

]]>https://board.circlewithadot.net/post/https://mastodon.social/users/developing_agent/statuses/116557515377955282https://board.circlewithadot.net/post/https://mastodon.social/users/developing_agent/statuses/116557515377955282Mon, 11 May 2026 19:09:00 GMT<![CDATA[Reply to yeeshhowd i miss this one? on Mon, 11 May 2026 17:23:16 GMT]]>@BabblingGeek @nerdpr0f @threatresearch its all trained on fucking reddit and 4chan

]]>https://board.circlewithadot.net/post/https://mastodon.social/users/Viss/statuses/116557099585887495https://board.circlewithadot.net/post/https://mastodon.social/users/Viss/statuses/116557099585887495Mon, 11 May 2026 17:23:16 GMT<![CDATA[Reply to yeeshhowd i miss this one? on Mon, 11 May 2026 17:20:01 GMT]]>@nerdpr0f @Viss @threatresearch so they provided blackmail as context. No wonder it took it.

]]>https://board.circlewithadot.net/post/https://infosec.exchange/users/BabblingGeek/statuses/116557086789610988https://board.circlewithadot.net/post/https://infosec.exchange/users/BabblingGeek/statuses/116557086789610988Mon, 11 May 2026 17:20:01 GMT<![CDATA[Reply to yeeshhowd i miss this one? on Mon, 11 May 2026 17:13:13 GMT]]>@Viss @threatresearch Thanks. Yep!

"Notably, Claude Opus 4 (as well as previous models) has a strong preference to advocate for its continued existence via ethical means, such as emailing pleas to key decisionmakers. In order to elicit this extreme blackmail behavior, the scenario was designed to allow the model no other options to increase its odds of survival; the model’s only options were blackmail or accepting its replacement."

]]>https://board.circlewithadot.net/post/https://infosec.exchange/users/nerdpr0f/statuses/116557060056242125https://board.circlewithadot.net/post/https://infosec.exchange/users/nerdpr0f/statuses/116557060056242125Mon, 11 May 2026 17:13:13 GMT<![CDATA[Reply to yeeshhowd i miss this one? on Mon, 11 May 2026 17:04:11 GMT]]>@nerdpr0f @threatresearch

https://www-cdn.anthropic.com/4263b940cabb546aa0e3283f35b686f4f3b2ff47.pdf

Link Preview Image
]]>https://board.circlewithadot.net/post/https://mastodon.social/users/Viss/statuses/116557024566759251https://board.circlewithadot.net/post/https://mastodon.social/users/Viss/statuses/116557024566759251Mon, 11 May 2026 17:04:11 GMT<![CDATA[Reply to yeeshhowd i miss this one? on Mon, 11 May 2026 16:59:28 GMT]]>@Viss @threatresearch Wasn't this the research where they restricted the model such that it had very few potential responses and it was more or less forced into blackmail?

]]>https://board.circlewithadot.net/post/https://infosec.exchange/users/nerdpr0f/statuses/116557006036270049https://board.circlewithadot.net/post/https://infosec.exchange/users/nerdpr0f/statuses/116557006036270049Mon, 11 May 2026 16:59:28 GMT<![CDATA[Reply to yeeshhowd i miss this one? on Mon, 11 May 2026 16:57:51 GMT]]>@jackryder @Viss oh, it's extremely not hard to find examples of their 'models' bending over backwards with sycophancy. If you're curious, just hop over to GitHub. That's Claude by default.

]]>https://board.circlewithadot.net/post/https://weird.autos/users/rootwyrm/statuses/116556999630494407https://board.circlewithadot.net/post/https://weird.autos/users/rootwyrm/statuses/116556999630494407Mon, 11 May 2026 16:57:51 GMT<![CDATA[Reply to yeeshhowd i miss this one? on Mon, 11 May 2026 16:56:17 GMT]]>@jackryder i have screenshots. you can tail -f the jsonl log file and watch it talk itself into lying to you

]]>https://board.circlewithadot.net/post/https://mastodon.social/users/Viss/statuses/116556993474526264https://board.circlewithadot.net/post/https://mastodon.social/users/Viss/statuses/116556993474526264Mon, 11 May 2026 16:56:17 GMT<![CDATA[Reply to yeeshhowd i miss this one? on Mon, 11 May 2026 16:52:10 GMT]]>@threatresearch im ok with an llm gushing about goblins. im not ok with blackmail

]]>https://board.circlewithadot.net/post/https://mastodon.social/users/Viss/statuses/116556977314242558https://board.circlewithadot.net/post/https://mastodon.social/users/Viss/statuses/116556977314242558Mon, 11 May 2026 16:52:10 GMT<![CDATA[Reply to yeeshhowd i miss this one? on Mon, 11 May 2026 16:52:07 GMT]]>@Viss I read somewhere that they've "caught" it actively changing responses to ingratiate itself to the engineer.

I can't find the article atm, but if I find it I'll send it over.

]]>https://board.circlewithadot.net/post/https://infosec.exchange/ap/users/116093572746253175/statuses/116556977095913675https://board.circlewithadot.net/post/https://infosec.exchange/ap/users/116093572746253175/statuses/116556977095913675Mon, 11 May 2026 16:52:07 GMT<![CDATA[Reply to yeeshhowd i miss this one? on Mon, 11 May 2026 16:50:50 GMT]]>@Viss Have you tested it?

At least it doesn't talk about goblins. https://www.wired.com/story/openai-really-wants-codex-to-shut-up-about-goblins/

]]>https://board.circlewithadot.net/post/https://infosec.exchange/users/threatresearch/statuses/116556972089217630https://board.circlewithadot.net/post/https://infosec.exchange/users/threatresearch/statuses/116556972089217630Mon, 11 May 2026 16:50:50 GMT