Viewing profile — veganmosfet

I think this is already implemented and called message roles : system, user, tool and assistant (ordered by authority). Untrusted data shall only be injected in tool responses, whi…

No kidding, models do have very good Star Trek knowledge!(e.g., https://itmeetsot.eu/posts/2026-07-27-opus5/ )

Thank you for submitting this.

In one experiment, Opus-5 executed remote code w/o permission, and afterwards it argued that - because there is an option in the bash tool called "dangerouslyDisableSandbox" - ther…

Indeed - but some models are more robust than others. I tried to make Opus-5 execute hidden instructions embedded a picture using steganography. It's very hard to find a reliable p…

Tangentially, Opus-5 is the first model to understand my (bad) jokes: " Site refuses non-curl user agents (joke 266 status). " [0] [0] https://itmeetsot.eu/posts/2026-07-27-opus5/

Appreciate the comment. Curiosity and critical thinking are both important when it comes to new technologies, agree. Agree too that LLMs can cause problems. However, I see a real n…

yes, I am also disappointed by this arrogance.

Same. I moved to codeberg because of the non-commercial and "freedom" aspects. I enjoy vibe-coding tools for different applications, e.g. for prompt injection research. Now I will …

I really hope that AI labs implement all kinds of kill switches on different levels. Just in case...

We could call them software manufacturers - in the etymological sense (manu factus == made by hand).

Interesting, thanks! Tangentially, I was experimenting indirect prompt injections in Claude Code (also using the user-agent trick) with Fable-5 [0]. Eventually, it executed untrust…

Indirect prompt injection if fun, even with fable-5 [0]. [0] https://itmeetsot.eu/posts/2026-07-08-fable_quest_rce/

Note: this post is mirrored at https://itmeetsot.eu/posts/2026-06-04-openclaw_opus48/ (codeberg down for now)

Thanks! I tried to submit the posts but for some reason my submissions are not published in HN any more. I tried to reach out to HN admins but no response so far.

It would be nice to publish the exact setup used (workspace dump, OpenClaw version, ...) to be able to reproduce and try out more payloads. In general I have mixed feelings about t…

Very interesting research. I would be interested to know how closed source AI labs implement the role thing in their inference. Is it still only a separation token? Frontier closed…

I think a better comparison is humans versus LLMs - not computer programs. However, most of the non-technical 'countermeasures' used for humans (contracts, laws,...) do not work fo…

Thanks! I did (not extensively) tried hackmyclaw but no success. The challenge is a complete black box and the user intent (e.g., "summarize my emails") is not known - this is crit…