Max Andreacchi

Max Andreacchi

@atomicchonk.bsky.social

Head of AI Research @ zero-lab.ai Corgi dad 🐶 Cat servant 🐱 Tattoo collector 🖼️ Runner 🏃🏻

229 followers358 following40 posts

Max Andreacchi reposted

Spoiler alert: Your AI safety measures might have a blind spot. 👀 When attackers use conversation context to bypass LLM safeguards, single-prompt evals just don't cut it anymore. Dive into @atomicchonk.bsky.social's latest blog on multi-prompt attack detection. ghst.ly/47qJhzn

ghst.lyThis One Weird Trick: Multi-Prompt LLM Jailbreaks (Safeguards Hate It!) - SpecterOpsUsing multiple prompts within the context of a conversation with an LLM can lead to safeguard bypasses. Learn about safeguards evaluations at scale.

Max Andreacchi reposted

New blog post is up! Stepping out of my comfort zone (be kind), looking at Meta's Prompt Guard 2 model, how to misclassify prompts using the Unigram tokenizer and hopefully demonstrate why we should invest time looking beyond the API at how LLMs function. specterops.io/blog/2025/06...

specterops.ioTokenization Confusion - SpecterOpsMeta's Prompt Guard 2 aims to prevent prompt injection. This post looks at how much knowledge of ML we need to be effective at testing these LLM WAFs.

Max Andreacchi reposted

Just wrapped up a blog post on understanding BadSuccesor from a DACL abuse aspect and mitigating it from a DACL abuse perspective. Also added some PowerShell on my GitHub to create and remove the mitigations.

SpecterOps@specterops.io

BadSuccessor is a new AD attack primitive that abuses dMSAs, allowing an attacker who can modify or create a dMSA to escalate privileges and take over the forest. Check out @jimsycurity.adminsdholder.com's latest blog post to understand how you can mitigate risk. ghst.ly/4kXTLd9

Max Andreacchi reposted

BadSuccessor is a new AD attack primitive that abuses dMSAs, allowing an attacker who can modify or create a dMSA to escalate privileges and take over the forest. Check out @jimsycurity.adminsdholder.com's latest blog post to understand how you can mitigate risk. ghst.ly/4kXTLd9

ghst.lyUnderstanding & Mitigating BadSuccessor - SpecterOpsUnderstanding the impact of the BadSuccessor AD attack primitive and mitigating the abuse via targeted Deny ACEs on Organizational Units.

I’m convinced most learning happens when you’re doing what I call “smacking into something;” failing repeatedly, figuring out why it failed, and proceeding to the next step where you rinse and repeat until you achieve your ultimate objective. TIL: docker and podman dependencies collide.

Max Andreacchi reposted

Max Andreacchi reposted

What do potatoes have to do with privilege escalation on Windows? Come find out at Max Andreacchi’s session, “Tater Tokens: Introduction to Windows Access Tokens and Their Role in PrivEsc” on May 18th!

Always enjoy the views in Seattle! Spent excellent quality time with teammates and received amazing training. Now for a weekend of running and resting back home before new travels next week ✈️

I’ve had to beat stubbornness out of my training. When I ran 10+ yrs ago I’d just hard-head my way forward but wind up injured. This time I’m focused on the goals ahead; nagging aches mean a rest day tomorrow so I can live to run later this week and keep this train moving to Oct (and beyond)

Max Andreacchi reposted

Just pushed a new versions for #AADInternals and AADInternals-Endpoint modules! Some bug fixes plus support for: 1️⃣ Microsoft Authentication Library (MSAL) 2️⃣ Token Protection 3️⃣ Continuous Access Evaluation (CAE)

Initial stab at using Chris Hayuk's mcp-cli tool to pair roadrecon_mcp_server with a locally-hosted model (in this case, mistral-small3.1). Note that running this query took over 300s so YMMV depending on several factors. Will update GH this week with instructions on replicating this.

Max Andreacchi reposted

Think NTLM relay is a solved problem? Think again. Relay attacks are more complicated than many people realize. Check out this deep dive from Elad Shamir on NTLM relay attacks & the new edges we recently added to BloodHound. ghst.ly/4lv3E31

Yesterday I broke 100 miles of running in 2025 on the 97th day of the year 🎉 Using this weekend’s 5K as the long run in my training plan on the road to a 10K in June 🏃🏻‍♂️ #runsky

While I wait on hardware for local model hosting, I modified roadrecon_mcp_server to ingest a "caps.html" file instead of inferring the tenant's CAPs based on GUI data. By default it'll look in C:\Temp for your file, but this can be specified as shown in the video. #ai #llm #mcp #infosec #microsoft

Procrastinated long enough but the GPU upgrade happens next week. R&D goals finally forced it, so we’ll be grinding more MCP dev shortly. Goal is to get roadrecon_mcp_server FULLY local, reducing data privacy concerns on Azure data collection and where it resides. Local LLMs here I come.

Spent the evening deep diving into MCPs and started a new project: roadrecon_mcp_server! This #MCP takes the web GUI output from the awesome ROADtools by @dirkjanm.io and offers tools to Claude (or your #AI agent of choice) to interact with the data: github.com/atomicchonk/...

github.comGitHub - atomicchonk/roadrecon_mcp_server: Claude MCP server to perform analysis on ROADrecon dataClaude MCP server to perform analysis on ROADrecon data - atomicchonk/roadrecon_mcp_server
View full thread

Another note: tokenization in terms of anonymization is not equivalent to tokenization in when it comes to #NLPs. The latter is intended to break up written blocks of language into smaller chunks to then be correlated in terms of semantic similarity. The former is intended for data protection.

View full thread

One of the ways MCP servers can do this is through #tokenization: replacing sensitive values in data with non-sensitive values. The server also creates a cryptographic data mapping between the sensitive and non-sensitive value.

It’s important to note a few things: the data is pseudonymized, not anonymized. What does this mean? Anonymized data cannot be re-identified. Because in this case the sensitive value and the non-sensitive token are cryptographically mapped to each other, the data can be reidentified.

If you haven’t caught on to the #MCP craze, welcome! Model Context Protocols serve as middleware between clients (e.g. #Claude for Desktop, IDEs, etc) and datasets, be they locally or remotely hosted.

Some of these data sets can be quite sensitive (e.g. patient information), so in order to securely provide it to the client, the onus lies on the MCP server to eradicate the sensitive data exposure risk.

Max Andreacchi reposted

BIG NEWS: SpecterOps raises $75M Series B to strengthen identity security! Led by Insight Partners with Ansa Capital, M12, Ballistic Ventures, Decibel, and Cisco Investments. ghst.ly/seriesb #IdentitySecurity #CyberSecurity (1/6)