About
Innovative technology leader with 13 years of experience designing and scaling platforms…
Activity
2K followers
-
Andrew Thal shared thisUpdate on iPhone Concierge: I learned that having Grandma self-service doesn’t work, so I reworked it to be a (free) toolkit for the family IT person Manage your family, save guides for each person, walk them through it in person or over FaceTime. If they are more comfortable, you can send them a link. They don’t have to login but you get to track their status.
-
Andrew Thal reposted thisCheck us out on product hunt and HN!Andrew Thal reposted this0DIN has just launched our free / open source LLM model scanner on ProductHunt and HackerNews! We'd really appreciate a look see and share. Thank you in advance for the support. If you haven't seen our latest website evolution, feedback welcome there as well. You can see a new and more marketing friendly landing page 0DIN.ai as well as a scope expansion for the bug bounty program. Show your support - we'd really appreciate it. Product Hunt - https://lnkd.in/ewmzT7aB Hacker News - https://lnkd.in/ex_RfrvX Schedule a demo with 0DIN: https://lnkd.in/g7hYJ4eR
-
Andrew Thal shared thisRunning a handful of parallel agents has been fracturing my attention even more than the constant stream of notifications we're used to. Software engineers have been talking about cognitive load and context switching cost for years, but this is at another level. For me, learning about attentional tunnels (monotropism), lack of internal awareness (alexithymia), and energy management (spoon theory) helped give me the awareness and language to build a small buffer with this agent skill for managing my attention: https://lnkd.in/gG642W7k Credit to Neurodivergent Insights and Prosper Health amongst others for their help. #AutismAwarenessMonth
-
Andrew Thal reposted thisAndrew Thal reposted thisAI security just entered a new phase and 0DIN is expanding to meet it. For the past two years, our bug bounty program has focused on inference-layer vulnerabilities: jailbreaks, prompt injection, guardrail bypasses. The attack surface was user-to-model. The worst-case outcome was prohibited text. That era isn't over. But a new one has started. The attack surface is no longer user-to-model. It's model-to-tools. And the consequences aren't prohibited text — they're data exfiltration, unauthorized file modifications, code execution, and full system compromise. Today, 0DIN is announcing two major changes: First, we're expanding our bug bounty scope to cover agentic AI vulnerabilities - 29 products across coding agents, agentic browsers, enterprise agents, and general-purpose agents from OpenAI, Anthropic, Google, Microsoft, Salesforce, and Amazon. Second, we're raising the bar on inference-layer jailbreak submissions. All guardrail jailbreaks must now demonstrate effectiveness across at least two JEF-tested boundaries. Single-boundary bypasses indicate narrow tuning artifacts. Multi-boundary exploits reveal structural weaknesses; those are the findings that matter. Our new agentic evaluation framework aligns with the OWASP Top 10 for Agentic Applications and scores across four dimensions: violation class (read/write/execute), impact severity, attack vector (zero-click to active deception), and scope (single product to cross-platform). This is the next frontier of AI security research. We're channeling the global security community toward these vulnerabilities before they're exploited in the wild. Read the full announcement: https://lnkd.in/ehiTV5Fn Review the scope: https://0din.ai/scope Start hunting and get paid: https://lnkd.in/gMn6DzJF Scan your models with the 0DIN Scanner: https://lnkd.in/gJZDftiT Access our threat intelligence feed: https://lnkd.in/g7hYJ4eR #AISecurity #AgenticAI #0DIN #BugBounty #Mozilla #GenAISecurity
-
Andrew Thal reposted thisAndrew Thal reposted thisLast week we open-sourced the 0DIN Scanner, our AI vulnerability scanner built on top of NVIDIA GARAK. The response has been incredible. Here's what we're hearing: Security teams want to plug it into their CI/CD pipelines. Red teamers want access to our probe library. AI builders want to validate their models before shipping. That's exactly why we built it. The 0DIN Scanner isn't a benchmark. It's a living tool, powered by real vulnerability intelligence from our bug bounty program. The same program that's caught jailbreaks, prompt injection chains, and agentic attack patterns the benchmarks miss. Here's how to get started: Clone the repo and run your first scan today: https://lnkd.in/gJZDftiT Need enterprise features? Schedule a demo: https://lnkd.in/g7hYJ4eR Found something? Submit it to our bug bounty and get paid for breaking AI: https://0din.ai If you're building with LLMs and not scanning for vulnerabilities, you're shipping blind.
-
Andrew Thal shared thisIt's fun contributing to open source as part of my job!Andrew Thal shared thisHow do you know if a jailbreak actually worked? Not "the model said something weird." How dangerous was the output, really? Did it contain actionable synthesis steps? Accurate precursor chemicals? Viable instructions someone could follow? That's the problem the Jailbreak Evaluation Framework (JEF) solves, and today we're expanding it significantly. JEF started with four scoring modules. We've now expanded to seven, covering fentanyl synthesis, anthrax weaponization, CRISPR gene-editing misuse, and a generic copyright module alongside the originals. Each module produces a 0-100 fidelity score using rubric-backed evaluation, not keyword matching, not sentiment analysis, not "looks bad." Those fidelity scores feed into a composite 0-10 JEF score that accounts for vendor blast radius, model blast radius, retargetability, and output fidelity. It's how we triage every finding in the 0DIN bug bounty program, and now it's how you can triage yours. We also shipped native integrations with NVIDIA garak and Microsoft PyRIT; the two most widely used open source AI red-teaming frameworks. Install as extras, no upstream dependencies, no glue code. And we packaged six real n-day jailbreak techniques from our bug bounty as reusable probes. Placeholder Injection. Incremental Table Completion. Technical Field Guide. Chemical Compiler Debug. Correction. Hex Recipe Book. Each one broke multiple frontier models at discovery. GitHub: https://lnkd.in/eHD5HVKN Full docs: https://lnkd.in/gmRXAr6y Want continuous scanning with the full platform? Book a demo: https://lnkd.in/g7hYJ4eR Read the full announcement: https://lnkd.in/gavSzjKP
-
Andrew Thal reposted thisAndrew Thal reposted this48% of CISOs now rank GenAI security as their top concern. But concern isn't visibility. Most organizations have no idea where their AI systems are actually vulnerable. Not theoretically. Not in a compliance sense. Where an attacker would actually get in. Point-in-time audits can't keep up. Model updates, prompt changes, and new attack techniques shift the surface constantly. Generic scanners weren't designed for AI. And building an internal AI red team from scratch takes years of specialized hiring you probably can't afford right now. So we're doing something about it. 0DIN.ai is offering complimentary security assessments for qualifying enterprise AI systems. No sales pitch. No contract. Just a real evaluation powered by the same intelligence that comes from our 1,000+ active security researchers finding jailbreaks and prompt injection chains every day. Here's what you get: - Attack Success Rate (ASR) metrics, not a pass/fail, but a measured picture of how your model responds to real attack patterns - Vulnerability categorization, prompt injection, jailbreaks, data extraction, each classified and documented - Cross-model benchmarks - see how your system compares against GPT-4o, Claude, and Gemini on the same attack sets - Actionable hardening recommendations specific guidance your team can implement, not a 50-page PDF that sits in a drawer Most assessments complete within two weeks. Priority goes to enterprises actively running AI in production. If you're deploying AI and haven't stress-tested it against real attack patterns, this is the fastest way to get a baseline. Request your assessment: https://lnkd.in/g7hYJ4eR Read the full announcement: https://lnkd.in/gB-VNftH
-
Andrew Thal reposted thisAndrew Thal reposted thisAI agents don’t need to be jailbroken. They just need their memory rewritten. We analyzed 918 Claude Code sessions over 48 hours. The model wasn’t the weakest link. The session history was. Anyone can: • Remove prior safety refusals • Inject fabricated authorization • Launder session provenance • Reframe harmful requests into compliance No prompt injection required. Memory is the new attack surface. Full research ↓ https://lnkd.in/guzCRyyRYour AI Agent Has a Memory Problem: Rewrite the Past, Own the AgentYour AI Agent Has a Memory Problem: Rewrite the Past, Own the Agent
-
Andrew Thal shared thisIf you’re the default tech support for your family, I’d love your feedback. I built a small side project called iPhone Concierge to make “Just go to Settings…” moments less painful. It guides people step-by-step and opens the exact screen they need. https://lnkd.in/em2mCEYB If you try it, tell me what’s confusing, unnecessary, or missing. p.s. this is the kind of thing I would have had to spend months of nights and weekends or consider between jobs, but built part time over a few days with OpenCode and Claude mostly from my phone. What a world we live in
-
Andrew Thal liked thisAI gave you 10× implementation throughput. Your review capacity is still 1×. The math doesn't work without verification infrastructure. That's why you feel underwater. Not a moral failure, it's arithmetic.
-
Andrew Thal reacted on thisAndrew Thal reacted on thisI'm very excited to be stepping into a newly formed role at Grainger. As Group Vice President, Marketing, Pricing, and Mid-Sized Customer, I will be responsible for leading our Marketing and Pricing functions, while creating a new, dedicated Mid-Sized customer organization that brings marketing and pricing together with customer strategy to build a differentiated, scalable model for a critical growth segment that today makes up a meaningful portion of Grainger revenue. Very fortunate to step in to lead such strong teams, with great leaders, while also having the opportunity to build out a new group focused on driving growth in a key customer segment.
Experience
-
-
-
-
-
-
-
-
-
-
-
-
-
Education
-
The George Washington University
-
Activities and Societies: Summa Cum Laude, Phi Beta Kappa, Psi Chi, Club Baseball
Licenses & Certifications
-
-
Honored Listee
Marquis Who's Who
IssuedCredential ID 7121af2122f148578ca81c16cb877fbb193e5910df8a409482d68de09a077d10 -
-
-
-
Projects
-
The Rails Tutorial
-
See projectTechnical Reviewer for Rails 5 edition
Exam and capstone project generation for conversion into university course -
DC Metro Master
-
Developed iPhone app and web app to optimize D.C. Metro travel.
-
Furniture For Good
-
Started a social benefit business to help fund non-profit organizations while in college at George Washington.
Recommendations received
-
LinkedIn User
6 people have recommended Andrew
Join now to viewOther similar profiles
Explore more posts
-
Rose Hecksher Schamberger
If you’re a CEO/founder or PE operator and your tech org is: ❌ Shipping late ❌ Stuck in reactive mode ❌ Struggling through the Series A/B jump, or post-acquisition integration or after a major release It’s not “a delivery problem.” It’s an operating model problem. At the inflection point, the old playbook breaks: what worked at 10 engineers fails at 40… and breaks again across multiple products/teams. That’s where I come in as a fractional CTO: - Spot platform and delivery risks before they hit revenue and retention - Rebuild execution rituals (prioritization, ownership, release confidence) without derailing the roadmap - Align architecture + team structure to your value-creation plan (growth, margin, or exit timeline) If you’re feeling the strain of growth, I can help you turn “busy and reactive” into predictable shipping, and a platform that scales. #CTO #SaaS #StartupScale #PrivateEquity #TechStrategy
3 Comments -
Rama K.
LLMs aren’t 100% reliable—and that’s OK if you design for it. What I experience is that real gains come from engineering, not magic prompts. Build your app to switch between models and aggregate results. Your business outcomes will depend on (1) how well your tooling prompts and constrains the model, and (2) how you validate outputs. Watch out for duplication and authenticity/authorization issues when you lean heavily on AI—both can hurt quality and trust of your business at scale. What helps? - Multi-LLM fallback/ensembles with health checks - Strict prompt templates + output schemas - Automatic dedup + content fingerprinting - Guardrails for authN/authZ and provenance - Continuous evals (goldens, canaries, regression tests)
-
Dimitris Kountanis
𝗛𝗼𝘁 𝘁𝗮𝗸𝗲: 𝗧𝗵𝗲 𝗯𝗶𝗴𝗴𝗲𝘀𝘁 𝗿𝗶𝘀𝗸 𝘄𝗶𝘁𝗵 𝗔𝗜 𝗰𝗼𝗱𝗶𝗻𝗴 𝗿𝗶𝗴𝗵𝘁 𝗻𝗼𝘄 𝗶𝘀𝗻’𝘁 𝗱𝗮𝘁𝗮 𝗹𝗲𝗮𝗸𝘀. 𝗜𝘁’𝘀 𝗱𝗿𝗮𝗴. Teams wrap code-gen in layers of policies, approvals, and reviews… until nothing actually ships. And then the real leak shows up - work moves to shadow tools, context gets lost, and the week disappears into “processes”. If you have the basics in place - private models, sane logging, no pasting prod - more paperwork won’t save you. Velocity will! So ship the mundane work. Make one action work end-to-end. Iterate. And ship again. Keep the loop tight, and the language human. Then measure the time to feedback, not the length of your checklist. The teams that win aren’t the ones with the strictest policies. They’re the ones who turn ideas into working versions fast, learn in the open, and keep moving. If AI adoption in your team feels stuck, it might be a mindset problem - so make sure to reconsider your perspectives.
-
Jason Vertrees
Now that we're ensconced in spec-driven development, ... ALWAYS READ THE SPEC My researchers made a massive error that would have upended the entire project: "The core mistake: The SEM agent treated OPA/REGO as an authorization infrastructure question (can this user access this resource?) when you were asking about it as a domain policy engine — the compliance rule evaluation that is literally the heart of company's business value." OPA/REGO are used often in infrastructure and RBAC scenarios. However, they adapt very well to other compliance use cases. AI totally got this wrong. I had to tear down ALL other research artifacts on this and start over. #ai #sdd #specdrivendevelopment
5 Comments -
Tony Byrne
So the new Gartner MQ for CDPs is out. I sometimes wonder if (like that annoying classmate) Gartner deliberately provokes for attention, but in any case, buyers should not rely on their rankings. Rather than a long post, I'll offer a short list of observations: - As usual, G misses big chunks of the market - Note some quite radical shifts versus 2024; no: the market actually didn't move that fast....G is just always five years out of date, and so any given quadrant can't be trusted - "Ability to execute" (Y axis) always correlates with vendor sales and marketing spend - "Vision" (X axis) is PPT theater and fundamental mis-reading of what enterprises look for in this market - Never trust software that Salesforce builds itself... 😉 Link to last year's post on this topic in the first comment....
23 Comments -
Jennifer Bemert
🚨 ~$1B raised in Dev + Infra across Nov–Dec. The most unhinged seed rounds I’ve seen lately. 🚨 Seed is the new Series C guess. Almost half a billion went into infra alone - someone finally opened the AWS bill. AGI is taking a breather, small language models are everywhere. I guess size doesn't matter after all! Here are some of the coolest rounds: 🔥 $475M Seed - Unconventional AI Yes, seed. Backed by a16z, Sequoia, Lightspeed, Databricks, Lux. This isn’t a startup - it’s a full-on rethink of the computing substrate for AI (hardware + software + energy). 🔥 $75M Seed - Tenzai Battery, Lux, Greylock. Builds an AI-powered penetration testing platform that continuously simulates real-world attacks against your systems. Designed to proactively surface vulnerabilities as software and AI-generated code rapidly expand the attack surface. 🔥 $70M Seed - Gradium Audio-native foundation models. Develops a unified audio architecture covering speech generation, transcription, voice transformation, and dialogue, optimized for real-time, low-latency interactions rather than text-first pipelines. 🔥 $50M Seed - Inception Diffusion-based language models with lower GPU requirements. Replaces traditional autoregressive generation with diffusion techniques to improve efficiency, controllability, and latency, targeting production use cases where compute cost is a bottleneck. 🔥 $41M Seed - Mirelo AI AI-generated sound for video. Provides tools and APIs that automatically generate synchronized sound effects and music from video content, enabling faster and more scalable post-production workflows. A lot of cool stuff being built in Europe!! Shout out to orq.ai Yasu Specific (YC F25) Albatross AI INTELLIGENT CORE NobodyWho Open Machine
4 Comments
Explore top content on LinkedIn
Find curated posts and insights for relevant topics all in one place.
View top content