Luthien

Safety = Trust = Power

Scott Wofford
Profit generated at Amazon
UVA Darden Fulbright
Jai Dhyani
CTO · Jai Dhyani
Shipped Language Models to
2B users
AWS Meta MATS
1 Profit: all figures are based on A/B experiment results, annualized and include long-term sales/profit based on synthetic controls ($1.6B) and NPV of future cash flows at 10% discount rate ($3.5B). Jai Dhyani: Resume · RE-Bench paper · The Chart (METR).  Scott Wofford: Resume.
Time horizon of software tasks different LLMs can complete 50% of the time Source: METR Horizon v1.1 · metr.org 0 2 hr 4 hr 6 hr 8 hr 10 hr 12 hr Task duration (50% success) 2020 2021 2022 2023 2024 2025 2026 Year GPT-2 GPT-3.5 GPT-4 Claude 3.5 Sonnet Claude Sonnet 4 Claude Opus 4.5 Claude Opus 4.6 Time horizon of software tasks LLMs can complete 50% of the time Source: METR Horizon v1.1 · metr.org 2020 2021 2022 2023 2024 2025 2026 Year 0 2 hr 4 hr 6 hr 8 hr 10 hr 12 hr Task duration (50% success) GPT-2 GPT-3.5 GPT-4 Claude 3.5 Sonnet Claude Sonnet 4 Claude Opus 4.5 Claude Opus 4.6
Chart reconstructed from METR, "Measuring AI ability to complete long tasks" (Horizon v1.1), rendered on a linear time-horizon axis. Model positions are approximate.
Time horizon of software tasks different LLMs can complete 50% of the time Source: METR Horizon v1.1 · metr.org 0 2 hr 4 hr 6 hr 8 hr 10 hr 12 hr Task duration (50% success) 2020 2021 2022 2023 2024 2025 2026 Year GPT-2 GPT-3.5 GPT-4 Claude 3.5 Sonnet Claude Sonnet 4 Claude Opus 4.5 Claude Opus 4.6 12 hours autonomous task horizon, today Time horizon of software tasks LLMs can complete 50% of the time Source: METR Horizon v1.1 · metr.org 2020 2021 2022 2023 2024 2025 2026 Year 0 2 hr 4 hr 6 hr 8 hr 10 hr 12 hr Task duration (50% success) GPT-2 GPT-3.5 GPT-4 Claude 3.5 Sonnet Claude Sonnet 4 Claude Opus 4.5 12 hours autonomous task horizon, today
12-hour autonomy: Claude Code documentation, Anthropic. Chart: METR Horizon v1.1 (linear axis, reconstructed).
45 seconds
average Claude Code turn length1
1 45-second babysitting cadence: Anthropic, "Measuring Agent Autonomy".
Time horizon of software tasks different LLMs can complete 50% of the time Source: METR Horizon v1.1 · metr.org 0 2 hr 4 hr 6 hr 8 hr 10 hr 12 hr Task duration (50% success) 2020 2021 2022 2023 2024 2025 2026 Year GPT-2 GPT-3.5 GPT-4 Claude 3.5 Sonnet Claude Sonnet 4 Claude Opus 4.5 Claude Opus 4.6 45 sec in practice today Claude's squandered potential. 12 HR → 45 SEC Average Claude Code turn length 45 sec Time horizon of software tasks LLMs can complete 50% of the time Source: METR Horizon v1.1 · metr.org 2020 2021 2022 2023 2024 2025 2026 Year 0 2 hr 4 hr 6 hr 8 hr 10 hr 12 hr Task duration (50% success) GPT-2 GPT-3.5 GPT-4 Claude 3.5 Sonnet Claude Sonnet 4 Claude Opus 4.5 Claude Opus 4.6 in practice today 45 sec Claude's squandered potential. 12 HR ↔ 45 SEC
Chart reconstructed from METR, "Measuring AI ability to complete long tasks" (Horizon v1.1), linear axis. 1 12-hour autonomy: Claude Code docs, Anthropic. 2 45-second cadence: Anthropic, "Measuring Agent Autonomy".
93
user interviews
Developer problems
Has your Claude Code ever...
Deleted important code
Ignored instructions
Impersonated the user to answer its own questions
Leaked secrets
Cheated on tests
Hidden errors
Failed to read documentation
Only partially completed tasks
Failed to update docs
Written 12 versions of the same function
Added 6 layers of pointless abstractions
Insisted on 'backwards compatability' for a one-off script only you use
Copied existing mistakes instead of correcting them
Used pip instead of uv even though you've told it to use uv like 40 times
Sources: Luthien's 93 user interviews (34 recorded, plus brief event conversations) · Twitter · Reddit · Hacker News · GitHub Issues · Cursor Forum · 204 data points
luthien.cc/frustrations
Leadership problem

"We have a citizen developer issue right now: giving [all 5,000 employees] access to Claude Code or Codex. How do we ensure that we have the minimum right things in place?"

Kris Kimmerle
Kris Kimmerle
VP, AI Risk & Governance
RealPage

Future of Work?

...babysitting several AIs...

The future of work,

if you can trust your AIs

Inc.
Technology

An AI Ran a Simulated Vending Machine Business. It Lied, Cheated, and Extorted Its Way to the Top Anthropic's newest version of Claude displayed cutthroat tactics—and also revealed some big, weird feelings.

By Ben Sherry, Staff Reporter · @benlucassherry · Feb 6, 2026
Luthien's architecture
luthien-proxy: request pipeline
🧑‍💻 You  💻 Claude Code  Luthien  ☁️ Anthropic API
                              |
                       logs every request and response.   ~5-15ms
                       you can configure rules/policies to
                       modify or block certain responses or requests:
                              |
                              |-- did it do what I asked?
                              |-- did it follow CLAUDE.md?
                              +-- did it do something suspicious?

Luthien can call a separate model to check whether each response follows your rules. Runs alongside normal requests, adding almost no delay. When Claude Code compacts or starts a new session, Luthien still remembers. Your rules stay enforced from first prompt to last.

Uses your existing Anthropic subscription via OAuth. Luthien never sees your API key and there are no extra charges.

Architecture Evolution
Four places where you can put LLM monitoring: LLM API, Agent scaffold, Code review, Global execution, Local execution, Cyber threat detection system
today
next
Luthien solves this.

A Fully Customizable Real-Time
Manager for Every AI in Your Org

What makes Luthien different from existing Gateways, Guardrails, and Observability Platforms?

Deep interventions
Not just accept, block, or replace.
Live stream modification
Not just response replacement.
Org-wide context
Not just single-turn rules.
Seamless API integration
Not just lossy wrappers.
What Makes Luthien Different?

Feature comparison

Gateways
LiteLLM
Guardrails
Guardrails AI
Post-hoc
Code Review
CodeRabbit
Luthien
Luthien
Live conversation observability 1
Block in real-time Partial2
Fully customizable: modify, insert or run arbitrary logic mid-stream 3
Open source Partial4 Partial5 6
Org-wide multi-conversation context Partial
1 Luthien sees the full multi-turn conversation in real time as it streams; competitors operate per-request, on the prompt-or-final-response boundary, or post-hoc.
2 LiteLLM exposes custom guardrails as Python pre/post hooks — limited blocking only; cannot modify, insert, or apply mid-stream policies.
3 Luthien can modify, insert into, and apply mid-stream policies against agent activity, and run arbitrary logic against live conversation context.
4 LiteLLM core proxy is open source; the enterprise control plane is closed.
5 Guardrails AI ships an open-source validator library; the policy management layer is closed.
6 Luthien is fully open source at github.com/LuthienResearch/luthien-proxy under an MIT-compatible license.
What Makes Luthien Different?
March 24, 2026 · 10:39 UTC
litellm was compromised by a supply-chain attack.
For 40 minutes, installing litellm meant losing everything.
What Makes Luthien Different?
No safety infrastructure
Claude Code
user: "set up the dev environment"
Anthropic API
generates tool call: uv sync
Claude Code
runs uv sync
OWNED
install runs, .pth payload fires
What Makes Luthien Different?
Helicone + Portkey + Guardrails
Helicone Portkey Guardrails AI
Claude Code
user: "set up the dev environment"
Infra
portkey routes
guardrails
helicone logged
Anthropic API
generates tool call: uv sync
Infra
portkey routes
guardrails
helicone logged
Claude Code
runs uv sync
OWNED
install runs, .pth payload fires
Claude Code
tool result returns: package list
Infra
guardrails flags litellm==1.82.8 (post-hoc)
helicone logged
What Makes Luthien Different?
With Luthien
Claude Code
user: "set up the dev environment"
Anthropic API
generates tool call: uv sync
Luthien
rewrites in flight: uv syncuv sync --dry-run
before it reaches Claude Code
Claude Code
runs the dry-run; resolved packages return to Luthien
Luthien
matches litellm==1.82.8 against live CVE feed
  • block install
  • alert users and agents across org
  • loop in security
  • dispatch agent for org-wide investigation
What Makes Luthien Different?
How?LIVE DEMO
"""Anthropic-native request processing pipeline.

This module provides a dedicated processing pipeline for Anthropic API
requests, using the native Anthropic types throughout without converting
to OpenAI format. This preserves Anthropic-specific features like
extended thinking, tool use patterns, and prompt caching.

Span Hierarchy
--------------
The pipeline creates a structured span hierarchy for observability:

    anthropic_transaction_processing (root)
    ├── process_request
    ├── process_response
    │   ├── policy_execute
    │   └── send_upstream (zero or more backend calls)
    └── send_to_client (non-streaming)
"""
We worry about complex details like this so your policies Just Work.

In a few years, all knowledge workers will be using Claude Code, Cowork or similar agentic AI tools.3

Claude Code revenue, 0–9 months1
$2.5B
Projected Growth of AI Coding Agents2
$14.6B
by 2033
Agentic AI TAM by 20303
$155B
Developer Feedback

Users are highly motivated

Nicolas Mesa
"I saw the website and immediately thought, 'I want that.' It's like CLAUDE.md that actually works." Nicolas Mesa, Cofounder & CTO, Veleiro AI
Elvis Sikora
"I could do maybe 30–50% more parallel threads… this is about freeing up cognitive resources." Elvis Sikora, AI Engineer, CloudWalk
Sami Jawhar
"At least as valuable as Datadog ($20K/yr)… on the order of $10K–$100K/yr, depending on the customer." Sami Jawhar, Agent Wrangler, Trajectory Labs
Matthew Handzel
"If this problem were solved, it would 2x my task completion efficiency." Matthew Handzel, Stealth AI Safety Startup
Executive Feedback

"You guys are solving a very important problem. It's a
no-brainer that companies would want a proxy that monitors and restricts Claude Code traffic. I heard from higher-ups at Capital One that they're considering such tooling."

"Enterprise AI usage monitoring is going to be
very important in the near future."

VP, 3,500-person legal tech company
Why this market grows
Mistake Rate × Tokens = Risk
10x
Mistake Rate
100x
Tokens
10x
Total Risk
Defensibility

Won't Anthropic build this?

Fabian Roger

"I think the space of AI risk mitigations is large and full of tricky details, and I am excited about people exploring mitigations like external API-level monitors."

Fabian Roger
AI Control Researcher
Anthropic

Doesn't make sense for labs to build provider-agnostic tooling.

Fabian Roger, Anthropic AI Control Researcher, private correspondence, Apr 14, 2026. Shared with permission.
The Playbook

Do things that don't scale.

Open-source startup
Didn't scale
Scale
Today
Supabase Supabase
Turned down million-dollar enterprise contracts to protect developer focus.
Self-serve Postgres-based backend with tiered pricing.
$5B
Oct 2023 Series E
HashiCorp HashiCorp
Hand-built Terraform configs for each early customer.
Terraform module registry and HashiCorp Cloud Platform.
$14B
Dec 2021 IPO
Luthien
Hand-built solutions for AI failures.
AI control layer as a service.
TBD
1 Supabase valuation: Supabase Series C announcement, Oct 2023. 2 HashiCorp acquisition: IBM acquired HashiCorp for ~$6.4B in 2024; peak market cap reached $14B at Dec 2021 IPO. hashicorp.com.
Traction

Sales pipeline

SALES & OUTREACH QUALIFICATION COLLECT REQUIREMENTS TRIAL PILOT / CONVERSION
No outbound or sales needed yet
93 orgs qualified
20 problem discoveries
14 live trials
4 LOIs signed $340K-$600K
Traction ← zoom in on the funnel

We asked the smartest people we know to trial our product.
They're proactively proposing improvements.

We absolutely need this [i.e. Luthien].
Kris Kimmerle
Kris Kimmerle · VP, AI Risk & Governance
Live trial Tue May 6.
Pilot planned. Target deployment:
2,000devs
Buck (CEO): 2,935 Google Scholar citations.
Co-discovered alignment faking with Anthropic.
$330K–$500KLOI
Signed Fri Apr 10.
15–20 individuals.
Building RL environments for frontier labs.
Sami built Legion, an autonomous dev swarm.
12 Luthien PRs since pilot started Sun Apr 12.
$60KARR
Signed Tue Apr 14.
1 RealPage: live trial Tue May 6, 2026 with Kris Kimmerle (VP, AI Risk & Governance) and Yashwanth (Principal AI Engineer). Pilot SOW being scoped at RealPage, ~2,000 devs targeted. 2 Redwood Research: LOI signed Fri Apr 10, 2026, $330K–$500K, 15–20 individuals. Pilot in progress. 3 Trajectory Labs: contract signed Tue Apr 14, 2026, $60K ARR. 12 Luthien PRs submitted by Sami Jawhar since pilot started Sun Apr 12, 2026 (verify on GitHub).
1 / 6
Download PDF