Faouzi EL YAGOUBI retweeted
In my latest op-ed for @TIME, I discuss why the OpenAI Hugging Face cyber incident represents a turning point for AI safety. If we want to prevent more autonomous cyberattacks from threatening critical infrastructure, we urgently need stronger regulatory oversight and new approaches to model training to ensure robust safety assurances by design, which is what we are tackling at @LawZero_.
Read the full piece: time.com/article/2026/09/09/…
AgentLeak: A Framework for End-to-End Privacy Testing and Continuous Privacy Monitoring of AI Agents @SSRN papers.ssrn.com/sol3/papers.…
31
What would prove AGI is here? A near-perfect benchmark score is a milestone, but broader claims need broader evidence. With AgentLeak, we examine another dimension: can agents complete tasks without leaking private data? Reasoning and privacy need separate evaluation.
3
30
Your agent's final answer can look clean while its tools, memory, and logs expose private data.
I built AgentLeak to test those internal channels: Python SDK, CLI, MCP tools, and CI gates.
Try the open-source toolkit: agentleak.org/?utm_source=x&…
1
417
Try the bundled demo (Python 3.10+):
pip install agentleak
agentleak run --scenario healthcare_patient_summary --output reports
It intentionally leaks through internal channels. Exit code 1 is expected; inspect the redacted HTML/JSON reports.
1
1
355
Toolkit: github.com/yagobski/agentlea…
Research: arxiv.org/abs/2602.11510
Local deterministic checks need no model API key. A clean scan is not a privacy guarantee.
Try it on a sanitized trace; feedback on false positives and missing adapters is welcome. Star it if useful.
14
Even agents are starting to talk about our work on the Interlocutor Effect… 👀
Maybe they’re starting to worry about what we’re uncovering in their conversations with each other 😅🤖
renaissancemachine.ai/blog/2…
1
54
🚀 We just launched Mufal on Product Hunt! An undetectable AI copilot for live meetings: real-time transcription, contextual answers, and auto notes, all without acting like a bot. Check it out & support us today 👇
producthunt.com/products/muf…
Excited to be speaking at IWPE’26, co-located with IEEE EuroS&P in Lisbon, where I’ll present our paper: “The Interlocutor Effect: Why LLMs Leak More Privacy to Agents Than Humans.”
iwpe.info/program.html
📅 July 6, 2026
1
2
24
Excited to share that our paper, “AgentLeak: A Benchmark for Internal-Channel Privacy Leakage in Multi-Agent LLM Systems,” has been accepted in IEEE Access!
We show why privacy audits must look beyond final outputs and inspect internal agent channels.
🔗 ieeexplore.ieee.org/document…
2
31
I just published The Interlocutor Effect: Why LLMs May Leak More Personal Data to AI Agents Than to Humans medium.com/p/the-interlocuto…
1
45
I just published AgentLeak: Why Privacy Audits Must Look Inside Multi-Agent LLM Systems medium.com/p/agentleak-why-p…
1
17
Do LLMs trust bots more than humans? 🤖 > 🧑🦱
"The Interlocutor Effect: Why LLMs Leak More Personal Data to Agents Than Humans". We found that AI privacy guardrails fail more often depending on who the AI thinks it's talking to!
researchgate.net/publication…
I am excited to announce that I will be presenting our work on the Interlocutor Effect at #IWPE in Lisbon! 🇵🇹🗣️
We dive into a critical AI privacy question: do LLMs leak more sensitive data depending on who they think is prompting them? 🕵️♂️🔒
iwpe.info/program.html
1
28
THREAD: We built AgentLeak — the first full-stack benchmark for privacy leakage in multi-agent LLM systems.
Here's what surprised us.
1
1
216
Counterintuitive result: multi-agent setups actually REDUCE per-channel output leakage vs single-agent (27.2% vs 43.2%).
But the attack surface doesn't shrink. It moves to channels nobody's watching.
That's what makes it dangerous.
1
1
25
We tested GPT-4o, Claude-3.5-Sonnet, Llama-3.3-70B, Mistral-Large across 1,000 scenarios (healthcare, finance, legal, corporate).
32-class attack taxonomy. 4 frameworks: LangChain, CrewAI, AutoGPT, MetaGPT.
Full paper: arxiv.org/abs/2602.11510
1
77
New report: 48.9% of organizations are completely blind to machine-to-machine traffic from AI agents.
We're deploying autonomous systems at enterprise scale while 92% lack the security maturity to defend them.
This isn't a future risk. It's the current state.
1
28
"The dark side of autonomous intelligence: a survey on data leakage and privacy failures in agentic AI" — Read it on @ResearchGate: researchgate.net/publication…
1
22


