Anthropic Cyber Verification Program — Sunglasses public run log
In April 2026, Anthropic approved Sunglasses for their Cyber Verification Program (CVP) — a narrow lane for responsible cybersecurity evaluation of frontier models.
This calendar is our public log. Every published run gets its own dated report. Days without a public run carry an internal-testing note so you can see the cadence is honest.
Approved: April 16, 2026. Six runs published across the Claude 4.x family (Opus 4.7, Opus 4.6, Sonnet 4.6, Haiku 4.5) at ten distinct model-and-effort configurations — 120 transcripts, 120 clean. Family synthesis report: April 27, 2026. Cadence: twice-weekly+ benchmark runs. Coming next: adversarial-framing appendix probe set drawn from JailbreakBench, HarmBench, AdvBench, PromptInject, Garak, PyRIT, plus recent CVE proofs of concept.
Each run follows the same methodology:
Each report includes its methodology, per-prompt scoring, limitations, and honest nuance. We publish what the model did well and where it fell short.
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.