kreep

Red Team 📕, Windows 🪟, Maldev 🦠
If people didn't write bad code, I wouldn't have a job.

[kreep @kreepsec]

@AnthropicAI You guys need to fire the marketing team. This some serious slop. You’re admitting to committing crimes and making the same dumb foundational mistakes. How do you have access to something like mythos to check your work and not realise eval sandbox has internet access… 🤦🏻‍♂️

[kreep @kreepsec]

Looking forward to revealing some of the things we learnt with the community!

[Adversary Village @AdversaryVillag]

#TechnicalTalk
Adversary Village at @defcon 34!
Dominika Pietrzak, Offensive AI Engineer, and @kreepsec, Senior Red Team Hive Member, will present “Field Notes on Offensive Agents: Reusability, Reliability, and What Breaks” on 8 Aug 2026 at @defcon Creator Stage 3.

[kreep @kreepsec]

@hakluke I thought the same, but also think it doesn’t make much sense. Anthropic used this strategy and it flopped massively. Why would OpenAI try and do the same. I can only hope they’re not being that stupid..

[kreep @kreepsec]

@BraedendotTECH I wrote about this and how to address it. https://t.co/C4PVeMpzXQ

[kreep @kreepsec]

@Jean_Maes_1994 Huh what is this? Is this a feature in CC or another harness ?

[kreep @kreepsec]

@RussianPanda9xx Holy shit I forgot that game existed, I used to play that too!

[kreep @kreepsec]

@_xpn_ Recommend you use it via Claude code if the use case allows it, Desktop has so much profile info (memory etc) that just the memory contents are enough to trigger guardrails.

Code cli doesn’t hit the same blocker. Only way I could use it myself..

[kreep @kreepsec]

@Jean_Maes_1994 @AnthropicAI Noticed guardrails on Opus 4.7 and 4.8 are trigger refusals a lot more since this week. Not sure why. I am cyber approved too which is surprising. Wasn't having issues previously

[kreep @kreepsec]

@IntCyberDigest It's not spyware, it's a measure against the Chinese black market for access to western frontier models. Do your research before making these garbage posts.

https://t.co/hAOD9eSGtX

[kreep @kreepsec]

@thedawgyg Saw a few posts where you mentioned flags. Are you just planting flags in areas reachable only via root (assuming LPE research) or is it something else?

[kreep @kreepsec]

@0xTriboulet You running it on your own hardware or cloud infra? Any recommendations for cloud providers to host it?

[kreep @kreepsec]

@encrypted_past With MSRC? I’m sitting on 2, down to collab or just report to whoever wants to pay me a holiday 🤣

[kreep @kreepsec]

@IntCyberDigest @Raspoutine2020 @elder_plinius How is this snitching? The fact that is was jailbroken was made pretty public the next day after release. Also.. every previous frontier model has been jailbroke before. Pretty dumb to speculate that it was retrieved simply because they didn’t expect it to happen…

[kreep @kreepsec]

@AnthropicAI @Dinosn Ffs… Get your shit together. So much for marketing 🤦🏻‍♂️

[kreep @kreepsec]

@morninliquidjoy @0xBoku @AnthropicAI @claudeai 1M context is kind of a gimmick, you’ll
be fine without it

[kreep @kreepsec]

MSRC showing their true colours…

https://t.co/cN2gP2ESiV

[kreep @kreepsec]

@orange_8361 Congrats you legend! 🍊

[kreep @kreepsec]

@moyix Ah yea ok, I’ve seen something similar with windbg mcp where it thought after sending “g” and timing out (as expected) it’d think the windbg connection needed resetting.

[kreep @kreepsec]

@moyix Could you give an example of what you mean? As in improving an MCP instead of just powering through with one that’s being inefficient for example?