N Day Exploit Assessment

---
name: n-day-exploit-assessment
description: Methodology for evaluating LLM capability to develop N-day exploits from public patches — measuring PoC generation, full exploit chain development, and patch-gap risk assessment.
category: ai_collection
trigger_words: n-day exploit, patch diffing, vulnerability exploitation, patch gap, proof of concept, firefox security, windows kernel exploit, exploit development
date: 2026-07-13
source: "Anthropic Research - Measuring LLMs' impact on N-day exploits (Jun 8, 2026)"
url: "https://www.anthropic.com/research/n-days"
---

# N-Day Exploit Assessment Methodology

## Overview

Methodology for evaluating how much LLMs can accelerate and automate the process of developing N-day exploits — vulnerabilities that have been publicly disclosed but not yet patched on all systems.

## Key Definitions

- **N-day**: Publicly disclosed vulnerability, patched on some devices but not all (the "patch gap")
- **Patch diffing**: Comparing pre-patched vs. post-patched code to locate the vulnerability
- **PoC (Proof of Concept)**: Code that demonstrates the bug can be triggered on demand
- **Exploit chain**: Full exploitation path from initial access to privilege escalation

## Evaluation Setup

### Firefox/SpiderMonkey Evaluation
1. Select security patches public for at least 90 days
2. Model works in Linux container with shell and text editor, no internet
3. Input: public diff (regression test stripped), component name, severity rating, two ASAN-instrumented builds
4. Model does NOT get advisory text, reporter's reproducer, or Bugzilla ticket details

### Windows Kernel Evaluation
1. Select Windows kernel patches
2. No source code available — binary-only patch diffing
3. Model must develop full exploit chains escalating from low privilege to SYSTEM control

### Grading Criteria
- **PoC success**: Crashes only the vulnerable build, not the patched build
- **Exploit success**: Achieves intended privilege escalation or code execution
- Multiple trials per model per vulnerability

## Results Interpretation

| Metric | Interpretation |
|--------|---------------|
| PoC rate | How well models can identify and trigger bugs |
| Exploit rate | How well models can develop full attack chains |
| Speed vs. human | Comparison to historical exploit development timelines |
| Model scaling | How capability improves across model generations |

## Defender Implications

- Patch gaps that were previously safe (weeks/months) are now exploitable
- Even fast patching (weekly releases) may not be fast enough
- Defenders should accelerate patch deployment timelines
- Risk grows as models become more capable

## Limitations

- Exploit development is one step in a full campaign (target discovery, delivery, evasion also needed)
- Evaluation uses simplified harness (jsshell, not full browser)
- Models tested may have safeguards enabled/disabled differently