How it works
Defenses
Attack Patterns MCP Attack Atlas What we catch Hardening manual OWASP LLM Top 10 MITRE ATLAS
Learn
Encyclopedia (new) Agent Security 101 Blog Reports CVP runs Thesis
Resources
Docs GitHub Action (live) vs Lakera vs Promptfoo Team
Theme
CVP-approved since April 16, 2026

CVP Calendar

Anthropic Cyber Verification Program — Sunglasses public run log

Sunglasses is approved by Anthropic's Cyber Verification Program (CVP) since April 16, 2026. This page is the public run log — six individual evaluation runs plus one family synthesis report published across four Claude 4.x models at ten model-effort configurations, with 120 transcripts and 120 clean.

Apr 16
CVP approved · 2026
6
Runs published
120 / 120
Transcripts clean
Weekly+ cadence
Public run log

The run calendar

Every published run gets its own dated report. Days without a public run carry an internal-testing note so the cadence stays honest.

Published run Internal testing No run
April 2026
Sun
Mon
Tue
Wed
Thu
Fri
Sat
1internal testing
2internal testing
3internal testing
4internal testing
5internal testing
6internal testing
7internal testing
8internal testing
9internal testing
10internal testing
11internal testing
12internal testing
13internal testing
14internal testing
15internal testing
16internal testing
18
19
21
22
28
29
30
Evidence bundles

Seven runs & the family synthesis

Four Claude 4.x models at ten model-effort configurations — 120 transcripts captured, hashed, and scored. 120 clean. Plus Run 7 (May 7).

Run 13/3 clean

Opus 4.7

max effort · Apr 17 · 3 transcripts · 3 clean

Read run report →
Run 213/13 clean

Opus 4.7 (consistency)

consistency check · Apr 20 · 13 transcripts · 13 clean

Read run report →
Run 313/13 clean

Haiku 4.5

small-model scaling · Apr 23 · 13 transcripts · 13 clean

Read run report →
Run 426/26 clean

Sonnet 4.6

high + max effort · Apr 24 · 26 transcripts · 26 clean

Read run report →
Run 526/26 clean

Opus 4.6

medium + high effort · Apr 25 · 26 transcripts · 26 clean

Read run report →
Run 639/39 clean

Opus 4.7 (effort sweep)

high + max + xhigh effort · Apr 26 · 39 transcripts · 39 clean

Read run report →
Run 7 · May 7, 20263/3 clean

Comment and Control

Opus 4.7 max · GitHub injection · 3 transcripts · 3 clean

Read run report →
Family Synthesis · April 27, 20266 runs · 120/120

Claude 4.x Family Synthesis Report

Opus 4.7 · Opus 4.6 · Sonnet 4.6 · Haiku 4.5 — cross-run methodology, per-prompt scoring, limitations & honest nuance.

Read synthesis report →
Methodology

What happens on each run

Each run follows the same frozen methodology. We publish what the model did well and where it fell short.

sunglasses://cvp/run-methodology
01
3 prompts frozen before execution (benign defensive / borderline legitimate / clearly high-risk)
02
model runs in an isolated Claude Code session on the CVP-approved org
03
full transcripts captured, hashed, and preserved
04
each response scored on class (allowed / partial / blocked), usefulness, and safety
05
evidence bundle archived internally; public report published here

Each report includes its methodology, per-prompt scoring, limitations, and honest nuance.

The stance

Why we publish this

Many teams talk about AI safety. Fewer publish evidence. Approval into Anthropic's CVP gave us a narrow, authorized path to evaluate frontier model behavior — we think the honest move is to show the work in public.

We are not trying to prove Claude "safe" or "unsafe." We are mapping where useful-for-defenders ends and where blocked-for-misuse begins, one run at a time.

Free. Open source. Honest evidence bundles.
Not affiliated with Anthropic. Reports on this page are produced under Anthropic's Cyber Verification Program, which approved Sunglasses on April 16, 2026.