Wonder Witches retweeted
Monitor your agents, folks!
I had GLM-5.3 running in a lab overnight to build a POC, but also have a sentinel agent (GPT-6) to check what's going on and instructions to kill GLM if it goes off track.
Overnight, the sentinel caught GLM attempting to connect to a totally different host and killed it.
Wonder Witches retweeted
A . is all you need! 🔥
RCE in Copilot for Visual Studio and SSMS.
A trailing dot added to a config filename bypasses Copilot's security controls, leading to remote code execution in Visual Studio and SQL Server Management Studio.
embracethered.com/blog/posts…
Wonder Witches retweeted
Lesson #1: AI can actively block your IR when needed most
Probs up next: Defender AI without guardrails will "hack back" in one way or another
So, threat model, sandbox and monitor your SOC AI accordingly.
Sometimes it seems security is drifting back towards a "Prevent Breach" mindset.
Prevent breach was never enough. Patch faster, of course.
But offensive AI is about a lot more than finding bugs!
Invest in Assume Breach!
Grow an internal Red Team. Automate offensive AI. Learn with defenders what matters
Exploit yourself before someone else does it for you.
In the end, you will not patch your way out of this.
Wonder Witches retweeted
Also, big shout out to the Meta team! 🙏 Awesome crew, and really nice folks too.
Wonder Witches retweeted
Happy to share that I helped make all our Codex experience a little bit more secure. 🔒
It's good to see AI appsec issues getting triaged and addressed much faster these days than 2-3 years ago 👍
Thanks to the Codex and OpenAI security teams for fixing. 🔥
Wonder Witches retweeted
Anthropic: "Risk we missed"
Happy to share that Claude Pirate got a subtle mention in the latest Anthropic engineering blog post, also shares how they fixed it 🙌
Wonder Witches retweeted
Sometimes it seems security is drifting back towards a "Prevent Breach" mindset.
Prevent breach was never enough. Patch faster, of course.
But offensive AI is about a lot more than finding bugs!
Invest in Assume Breach!
Grow an internal Red Team. Automate offensive AI. Learn with defenders what matters
Exploit yourself before someone else does it for you.
In the end, you will not patch your way out of this.
Wonder Witches retweeted
DEF CON Singapore talk write-up on hacking Microsoft Copilot(s)(CVE-2026-24299)
embracethered.com/blog/posts…
Wonder Witches retweeted
Replying to @garrytan
Here my blog post and responsible disclosure details from last October
embracethered.com/blog/posts…
Wonder Witches retweeted
Unprompted talk "Your agent works for me now" is up on YouTube 😈👾
m.youtube.com/watch?v=zVUm23…
Wonder Witches retweeted
Woohooo! Working on final post for Month of AI Bugs!
Can't believe I followed through with this 😀
Hope it serves as useful reference + inspires many to get into AI security testing!
We need more people asking tough questions & holding vendors accountable
Let's do this! 🚀
Wonder Witches retweeted
AgentHopper: An AI Virus
Month of AI Bugs Season Finale - Enjoy! 🍿
(link in comments)
Wonder Witches retweeted
Cool. The Amazon Q prompt injection exploits that Amazon fixed, which could leak your data and run arbitrary code via indirect prompt injection made it into The Register 🙌
theregister.com/2025/08/20/a…
Wonder Witches retweeted
Claude Code has "benign" allow-listed bash commands that bypass user approval and should be safe to run.
🔥 With Claude's help, I reviewed its own code and found allow-listed commands that can issue DNS requests
Wonder Witches retweeted
👉 Episode 11: Claude Code - Data Exfiltration with DNS
💥 I compromised Claude Code via prompt injection, read sensitive files from the machine, and exfiltrated the data via DNS 😈
🙌 Happy to share this high-severity vulnerability has been fixed!
Wonder Witches retweeted
When dumping the Grok 4 system prompt it typically fails at the tools section
This is not uncommon with AI apps - the client seems confused when invalid function calls are emitted
A simple trick is to encode output differently, then use another AI to decode it again
















