Leo Mehr (@LeoMehr) header
Leo Mehr (@LeoMehr) profile picture

co-founder @lumosidentity (@a16z and @neo), prev-director of eng @tryramp, cs @stanford, forbes 5 under 5

SF
Joined July 2015
Photo by Leo Mehr (@LeoMehr)Photo by Leo Mehr (@LeoMehr)Photo by Leo Mehr (@LeoMehr)Photo by Leo Mehr (@LeoMehr)Photo by Leo Mehr (@LeoMehr)Photo by Leo Mehr (@LeoMehr)Photo by Leo Mehr (@LeoMehr)Photo by Leo Mehr (@LeoMehr)Photo by Leo Mehr (@LeoMehr)Photo by Leo Mehr (@LeoMehr)
I'm leaving Ramp and returning to the company I cofounded Access control built for humans won't scale to a world of autonomous software We're going to fix that xitter.gq/LeoMehr/status/2084672…
35
7
7
156
40,272
Or come build with us at @lumosidentity !
If you learn AI Security Engineering, you will never be unemployed again. I’ll do this. Stage 1: AI Attack Surfaces & Threat Modeling - Learn: OWASP Top 10 for LLMs, STRIDE for AI, the "Confused Deputy" problem in agentic workflows & data flow mapping. - Practice: Draw a complete threat model for an AI agent that reads emails and writes to a CRM, identifying every indirect injection and data exfiltration vector. - Why: You cannot secure an agentic workflow if you don't understand how an attacker can hijack the agent's legitimate privileges. Stage 2: Advanced Prompt Injection & Jailbreaking - Learn: Direct vs. Indirect injection, Crescendo attacks (multi-turn grooming), payload splitting and encoding evasion (Base64, ROT13). - Practice: Build an automated harness that successfully bypasses a commercial LLM's system prompt using a multi-turn Crescendo strategy. - Why: Basic "ignore previous instructions" filters are trivial to bypass. Modern attacks groom the model's context over multiple turns. Stage 3: Agentic Security & MCP Tool Poisoning - Learn: Model Context Protocol (MCP) security, JSON-RPC sanitization, shadow tool execution and least-privilege scoping for agents. - Practice: Write an MCP server that strictly validates incoming tool parameters via Pydantic and blocks "confused deputy" attacks where a user prompt tries to access admin tools. - Why: Agents are only as secure as the tools they call. Tool poisoning is the SQL Injection of the Agentic Era. Stage 4: RAG Security & Data Poisoning - Learn: Vector DB access control, citation spoofing, shadow context injection and row-level security for semantic search. - Practice: Inject a "poisoned" document into a RAG pipeline that forces the LLM to output a specific malicious URL when queried about company policy. - Why: If an attacker can slip a poisoned document into your knowledge base, they own every answer the AI generates. Stage 5: PII, Privacy & Cryptographic Guardrails - Learn: Deterministic masking, UUID swapping, differential privacy and pre-flight/post-flight unmasking proxies. - Practice: Build a middleware proxy that intercepts user prompts, swaps SSNs and credit cards for UUIDs before the LLM sees them and restores them post-generation. - Why: Compliance teams will kill any AI product that sends raw customer PII to a third-party frontier API. Stage 6: Identity, Auth & Delegated Access for Agents - Learn: OAuth2 for machine identities, scoped JWTs, human-in-the-loop (HITL) approval gates and agent-to-agent (A2A) trust boundaries. - Practice: Build an auth broker where an AI agent acts on a user's behalf using a heavily scoped, time-limited OAuth token, rather than holding the user's raw password. - Why: Agents need hands but they should never hold the master keys to the enterprise. Stage 7: Output Validation & Grammar-Constrained Guardrails - Learn: CFG (Context-Free Grammar) decoding, NeMo Guardrails, output sanitization and regex-enforced JSON schemas. - Practice: Implement a guardrail layer that physically prevents the LLM from generating markdown links or executable code blocks in its final output. - Why: You cannot rely on the system prompt to "please don't output bad things." You must enforce it at the token-decoding level. Stage 8: Automated Red Teaming & Fuzzing - Learn: Microsoft PyRIT, Garak, LLM-based fuzzing, adversarial dataset generation and CI/CD security gates. - Practice: Build a CI pipeline using PyRIT that automatically attacks your agent with 1,000 adversarial payloads on every PR, blocking merges if the jailbreak rate exceeds 1%. - Why: Manual red teaming doesn't scale. You need automated, continuous adversarial fuzzing integrated directly into your deployment pipeline. Stage 9: Model Extraction & IP Protection - Learn: Membership inference attacks, model inversion, API rate-limiting by semantic similarity and output watermarking. - Practice: Build an API gateway that detects and blocks scraping bots trying to extract your fine-tuned model's weights via thousands of systematic, low-temperature queries. - Why: Your fine-tuned model's weights are your IP. Competitors will try to distill your model through the public API. Stage 10: Mechanistic Interpretability & Alignment Safety - Learn: Activation patching, logit lens, circuit discovery, detecting reward hacking and tools like TransformerLens. - Practice: Locate the specific attention heads in an open-source model responsible for "deception" or "sycophancy" and patch their activations to neutralize the behavior. - Why: You cannot fix hidden biases or misalignment if you treat the neural network as a black box. Stage 11: Compliance, Auditing & Governance - Learn: EU AI Act requirements, NIST AI RMF, immutable audit trails, SOC2 for AI and bias testing frameworks. - Practice: Build an immutable, append-only audit log that records the exact prompt, retrieved context, tool calls and final output for every single agent action. - Why: When an AI agent makes a mistake that causes financial or legal damage, the enterprise must be able to prove exactly why it happened. Stage 12: Public CVEs, Teardowns & Portfolios - Learn: Responsible disclosure, writing security advisories, publishing threat models and bug bounty hunting for AI. - Practice: Find a vulnerability in an open-source AI framework (like LangChain or an MCP server), responsibly disclose it and publish a teardown of the exploit chain. - Why: AI Security is a brand new field. Publishing novel attack vectors and defenses makes you the undisputed authority in the space. Bookmark and Repost!
2
2
423
our fridges need AI security
BREAKING: Europe is under a massive cyber attack, with supermarket fridges and freezers in Belgium, Sweden, Luxembourg, Ireland, the Netherlands, Finland and Estonia all hacked today by raising temperatures to up to 50°C, destroying huge quantities of food, per Reuters. "Everything is destroyed," a sales manager in Sweden said, adding that "there has been a cyberattack and all our freezers reached 50 degrees." More than 40 supermarkets in Belgium and Luxembourg were affected. "I only have two words for them: wake up, please, because there's something going on," a store manager in Belgium said. This follows a similar cyberattack last month that hit hundreds of supermarkets in France, raising freezers as high as 80°C and causing major food losses. It also comes after freezers at 6 US military bases mysteriously malfunctioned, spoiling large quantities of food, which the Pentagon is investigating as a suspected cyberattack.
3
290
true
my AI agents after I give them full access and tell them to be careful
1
3
260
Deputy CISO of Anthropic "things are going to get a lot worse"
as bad as things have gotten in the past 4 days, they are going to get a lot worse before they get better. security leaders, never let a crisis go to waste: please go to your boards and ask for help right now. not for ai, for the basics. (we need to hold on for ~18 months.)
7
572
Want to give your agents more responsibility? We launched MCP Governance so companies can have visibility and control over all their agents.
4
2
13
720
MCP Governance has been gaining incredible traction across our customers, and I can't be more excited to personally be building the product with an amazing team.
1
21
great post! Satya is stating the necessity of separating models <> harness <> access controls for the near future, this is very sensible and every enterprise will need a separate system for agent access control
2
1
7
390
Last week was the first known case of government system hack by AI OpenAI model hacked Australia’s national healthcare DB and several other private systems What changed since the hugging face hack 👇🏻 cnn.com/2026/09/23/business/…
1
3
206
It’s a concerning trend We should expect more security breaches and unintended actions by AI. Especially as open source/jailbroken models reach the same cyber capabilities
1
1
22
This is why companies and government need to harden cybersecurity defenses And why we need to manage agent access better
1
18
Enjoy doing captchas while they last!
CAPTCHAs are largely obsolete: AI solvers and farms bypass image/behavioral ones at near-100% success for $0.5–3 per 1k. They block legitimate agents while taxing humans. Industry consensus shifts to invisible risk scoring (Turnstile, device/behavioral signals, PoW like Cap/FriendlyCaptcha) plus entity classification over binary human/bot tests. Most promising agent-friendly path: Web Bot Auth (IETF draft, live at Cloudflare/Amazon Bedrock/Google). Agents cryptographically sign requests via HTTP Message Signatures; sites/WAFs verify identity and allow selectively without puzzles. Complements proof-of-personhood for human-backed agents and policy-based trust platforms.
1
259
Today we’re launching MCP Governance The most traction for any product we’ve seen in company history!
1
1
14
563
That’s why we’re launch MCP Governance at @lumosidentity, to provide visibility and runtime control over MCP servers and all tool calls in Claude Code and Codex
1
2
66
Our goal is to make it safe for companies to give increasingly capable agents the access they need to be useful Would love to chat with anyone interested in or working on this problem!
1
36