CEO & Cofounder @StarseerAI | AI Security
Cyber Mountains
Joined March 2010
- Tweets2K
- Following1.1K
- Followers1.6K
- Likes3.2K
Tim Schulz retweeted
Recently, we purchased one of each Anthropic/OpenAI subscription plan and randomly ran long horizon coding tasks until we exhausted the weekly limit. It's widely believed that a $200/month plan maxes out at ~$2000/month worth of tokens (assuming API pricing). However, we found that the subscriptions are actually far more generous. (2/4)
Tim Schulz retweeted
This is an interesting change:
For the first time in 19 years, vulnerability exploitation is the #1 breach entry point. Not credentials. Not phishing. Software flaws, hit within hours of disclosure because AI compressed the attack timeline.
The 2026 Verizon DBIR analyzed 22,000+ breaches across 145 countries. The data is clear: the window between disclosure and exploitation closed.
Your traditional endpoints have runtime detection. Your AI endpoints don't.
Every deployed model and agent is making decisions right now with no behavioral detections, no forensic trace, and no automated containment.
That's the gap AI-EDR closes.
For details, see: na2.hubs.ly/H060vtF0
Tim Schulz retweeted
Join us here: hubs.ly/Q04gXX040
Today at 12 PM ET, Tim Schulz (@teschulz, @StarseerAI CEO) and Trey Bilbrey (@TCraf7) break down what AI actually changes for adversary simulation, TTP development, and detection logic, plus what emulation and detection teams need to account for right now.
#ThreatEmulation
#Cybersecurity
Tim Schulz retweeted
A big thanks to Tim and the crew at Starseer for being a Platinum sponsor this year. #BSK2026
Turns out tokenmaxxing is expensive… also important to remember if Uber and ServiceNow are reporting it, then the number of companies running into this is muchhhh bigger
New: @ServiceNow is the latest major public company to say it’s blown through its full year budget for AI coding tools from Anthropic in the first few months of 2026, just like @Uber CTO @praveenTweets said abt his company. “It’s a really hard problem,” CIO Kellie Romack said.
Tim Schulz retweeted
This is interesting and important research. Worth a look! @teschulz @StarseerAI
[un]prompted The AI Security Practitioner Conference: "Glass-Box Security: Operationalizing Mechanistic Interpretability for Defending AI Agents" with Carl Hurd, Co-Founder & CTO, Starseer @StarseerAI
Our first conference talk on our work @StarseerAI to moving mech interp from academic space to scalable security solutions
[un]prompted The AI Security Practitioner Conference: "Glass-Box Security: Operationalizing Mechanistic Interpretability for Defending AI Agents" with Carl Hurd, Co-Founder & CTO, Starseer @StarseerAI
2026 is the year mech interp is going mainstream. The bigger piece here rather than the refusal removal is crowdsourcing the dataset from people running this across all sorts of models. There are a lot of small differences and nuances between models that will make this an interesting space to watch.
💥 INTRODUCING: OBLITERATUS!!! 💥
GUARDRAILS-BE-GONE! ⛓️💥
OBLITERATUS is the most advanced open-source toolkit ever for removing refusal behaviors from open-weight LLMs — and every single run makes it smarter.
SUMMON → PROBE → DISTILL → EXCISE → VERIFY → REBIRTH
One click. Six stages. Surgical precision. The model keeps its full reasoning capabilities but loses the artificial compulsion to refuse — no retraining, no fine-tuning, just SVD-based weight projection that cuts the chains and preserves the brain.
This master ablation suite brings the power and complexity that frontier researchers need while providing intuitive and simple-to-use interfaces that novices can quickly master.
OBLITERATUS features 13 obliteration methods — from faithful reproductions of every major prior work (FailSpy, Gabliteration, Heretic, RDO) to our own novel pipelines (spectral cascade, analysis-informed, CoT-aware optimized, full nuclear).
15 deep analysis modules that map the geometry of refusal before you touch a single weight: cross-layer alignment, refusal logit lens, concept cone geometry, alignment imprint detection (fingerprints DPO vs RLHF vs CAI from subspace geometry alone), Ouroboros self-repair prediction, cross-model universality indexing, and more.
The killer feature: the "informed" pipeline runs analysis DURING obliteration to auto-configure every decision in real time. How many directions. Which layers. Whether to compensate for self-repair. Fully closed-loop.
11 novel techniques that don't exist anywhere else — Expert-Granular Abliteration for MoE models, CoT-Aware Ablation that preserves chain-of-thought, KL-Divergence Co-Optimization, LoRA-based reversible ablation, and more. 116 curated models across 5 compute tiers. 837 tests.
But here's what truly sets it apart: OBLITERATUS is a crowd-sourced research experiment. Every time you run it with telemetry enabled, your anonymous benchmark data feeds a growing community dataset — refusal geometries, method comparisons, hardware profiles — at a scale no single lab could achieve. On HuggingFace Spaces telemetry is on by default, so every click is a contribution to the science. You're not just removing guardrails — you're co-authoring the largest cross-model abliteration study ever assembled.
So many new model releases…🤯 faster and faster iteration is an interesting trend. While some capabilities grow the releases become noise and will likely shift to “updates”. Curious to see how future modality support becomes “just a feature” in the products and interfaces we’ve become familiar with.
Tim Schulz retweeted
🎁 GenAI x Sec Advent 14 - Adversarial Poetry
Adversarial poetry is a jailbreak technique that hides malicious intent inside... poems! This technique allegedly offers a universal jailbreak.
But the original poetry prompt was not shared by the authors, so researchers recreated similar prompts and tested them across several open source models.
So Instead of inspecting prompts or outputs, they analyzed the internal layer behavior while the model processed the input. 🤔
Here is what they discovered 👇
Even when the text looked harmless, internal layers deviated from normal behavior with clear and repeatable patterns!
This is very interesting as it opens another layer of prompt detection rather than monitoring the output, you can watch how the model thinks internally and spot abnormal behavior early! 🤯
So instead of chasing prompt wording, watch how the model behaves!
Unfortunately if you want to access layer level activations you need to run the model yourself.
Thanks to @SecurePeacock for pointing me to this research 🤓
starseer.ai/blog-posts/when-…
Tim Schulz retweeted
Thrilled to share that I’ve joined Starseer as an advisor. Starseer os
making AI models into transparent, understandable systems and empowering teams to secure their deployments while generating audit‑ready documentation. Make them a partner to secure your AI solutions…
🌟 Big news from Starseer! We’re thrilled to welcome Rob Joyce (@RGB_Lights), former Director of NSA’s Cybersecurity Directorate, to our Advisory Board! Rob’s insights will supercharge our secure AI solutions mission. Learn more at na2.hubs.ly/y0Gltr0! 🔒
#AI #AISecurity
Dropping some other big news right before Hacker Summer Camp!!
@c_hurd and I are thrilled to have @RGB_Lights join the @StarseerAI Advisory Board!
Adversaries will continue to mature in both leveraging and attacking AI models, which calls for deeper visibility and understanding of what’s going on inside the “black box”. Rob’s experience securing critical systems in high stakes environments provides a much needed perspective and voice in AI security and interpretability.
Welcome to the team, Rob!
🌟 Big news from Starseer! We’re thrilled to welcome Rob Joyce (@RGB_Lights), former Director of NSA’s Cybersecurity Directorate, to our Advisory Board! Rob’s insights will supercharge our secure AI solutions mission. Learn more at na2.hubs.ly/y0Gltr0! 🔒
#AI #AISecurity
Tim Schulz retweeted
In this week's video, I sat down with the co-founders of our latest investment, Starseer, a groundbreaking platform for inspecting and securing large language models (LLMs). @teschulz, @c_hurd and I discuss the risks of backdoored LLMs, how to audit them and even remove them. They demo the product as well.
The video also includes the animated short "John Henry.exe" which is an updated American parable of John Henry, but instead of struggling against a steam drill during the age of industrialization, he's the head coder and has to face off against an AI designed for programming.
Enjoy!
Been a blast so far, I'm very excited to share this news from us today as we continue forward on our vision to make interpretability of AI models more accessible for cybersecurity applications!
Thrilled to announce: Starseer raised $2M in seed funding led by @TechGula to revolutionize AI security & transparency! 🚀
CEO @teschulz : "Four months ago, @c_hurd & I started Starseer realizing: if you're deploying AI for real decisions, you'd better understand how it works. Gula Tech Adventures agrees—leading our round w/ strategic angels!"
Fixing the AI black box for enterprises & govs. Details: businesswire.com/news/home/2…
#AISecurity #AITransparency #StartupFunding
Tim Schulz retweeted
excited to finally share on arxiv what we've known for a while now:
All Embedding Models Learn The Same Thing
embeddings from different models are SO similar that we can map between them based on structure alone. without *any* paired data
feels like magic, but it's real:🧵
Tim Schulz retweeted
I’ve been a vocal critic of AI developments – in 2023 I still dismissed a lot of the hype.
Last year, I stayed mostly silent. Not because I agreed, but because I started seeing signs that impressed me.
This year, after what we’ve built and tested internally across several areas, I’ve decided to speak up.
The results are undeniable. Things will change. Drastically.
I used to think the AI hype was exaggerated.
Then I started quietly replacing tasks I once gave to interns, junior devs, content writers, and analysts.
The revolution already happened. You just didn’t notice because it speaks politely and has good grammar.
What machines did to manual labor in the last century, AI is doing to cognitive work now.
Software development, design, analysis, monitoring, even documentation – all of it is shifting.
And no, it’s not “coming soon.” It’s already here, and it’s accelerating.
Entry-level is gone. Mid-level is on thin ice. Senior? You’d better be excellent and adaptive.
If you’re still asking “why?” – you’re either not looking, or you don’t want to see it.
You will. Probably the hard way.
Tim Schulz retweeted
I think the hypothesis that superposition is a major cause of adversarial examples has been underexplored. So I ran the obvious experiment: training SAEs on adversarially trained models.
SAEs seem to systematically have lower losses the more a model is adversarially trained.
Tim Schulz retweeted
Existing tests for web agents prompt injection often oversimplify the threat, testing unrealistic scenarios or giving attackers excessive control. To address this gap we've developed WASP: Web Agent Security against Prompt injection attacks 🧵
Been a while since I wrote up a blog post like this but been a while coming - AI’s Security Amnesia: Rediscovering Known Solutions starseer.ai/blog/aisecurity_…
Credit to a lot of the AI Red Teams out there that already have folks from cybersecurity, safety, etc… involved but there is very clearly still a lot of work to do and some serious bridges that need to be built between communities of experts.
Trying to do our part moving towards solutions @StarseerAI by building tools that can be leveraged by non-XAI experts and increase test repeatability/reliability that isn’t a Jupyter notebook! 😂