Daily News · 22 November 2026

Riemann Proof Certified, Inference Leaks & Upload2Shell

Interactive explorer for the Top 3 stories: formal Lean certification, canary test failures, and Ghostscript RCE.

Lean Library Size
~17M lines about
Released 20 Nov 2026 by Levent Alpöge
Certification Date
21 Nov 2026
Lean FRO & Mathlib Initiative endorsed
Formal Checks
4 checks Passed
2 comparison tools · 2 sound checkers
Human Review
Not started Open
No human understands it yet; Clay prize unstarted

Formal Certification Pipeline vs. Human Understanding

Toggle the 4 verification gates on the right to see why "certified" means logically valid in Lean, not yet understood.

Certified in Lean
[OK] All 4 formal checks pass: Lean & Mathlib communities have full confidence the proof is logically valid in Lean. Human understanding, peer review, and the Clay prize process have not started.

Verification Controls

Test what happens if any check is missing

MIT Scope Note (Andrew Sutherland)

Scale Explorer: About 17 Million Lines vs. Human Reading Speed

Adjust the user assumption sliders to see why machine checking replaces months of manual review.

Sliders = Assumption
500
10
Providers Off-Boarded
3 providers
By OpenRouter (22 Nov) · Separate notices: Novita, Phala, Together
ZDR Endpoints Cut
125 endpoints
Removed from OpenRouter ZDR eligibility
Together Leak Scope
3 models · 2 mo
June & July stored prompt samples (21 Nov)
Dragos Canaries Hit
5 canaries
Planted secrets found in leaked prompts

Together AI Prompt Leak Chain & Recommended Advice

How an AI coding agent commit led to prompt theft. Only ZDR affects prompt storage; rotate secrets and vet providers as general advice.

High Exposure
[!] Existing account without ZDR enabled: prompts from June & July could be stored and harvested if a provider credential leaks.

Account & Advice Checklist

Post recommendations (general advice)

Together AI Account Type
20

OpenRouter Action & Reported Incident Numbers

OpenRouter itself was not breached and notified affected customers; it separately lists public notices from Novita, Phala, and Together.

22 Nov 2026
CVE Identifier
CVE-2026-125396
Write-up by pwn.ai (21 Nov 2026)
Flaw & Sandbox
Use-after-free
Type 1 font loader escapes -dSAFER
Fixed Version
GS 10.08.0 Patched
Confirmed on oss-security (21 Nov)
Distro Lag
Uneven Check
Debian Bookworm vulnerable on 12 Nov

Upload2Shell Attack Path & Defense Simulator

Select one of the 3 remote entry vectors demonstrated by pwn.ai and test which defenses stop code execution.

RCE Reachable
[!] Vulnerable path open: viewing an email in Roundcube triggers thumbnail generation, reaching Ghostscript's Type 1 font loader and escaping -dSAFER.

Entry Vector & Defenses

Found by an AI model + human researcher

Remote Entry Vector

Live Mitigation Matrix Across All 3 Demonstrated Vectors

Patching Ghostscript (10.08.0) blocks all 3 vectors; CUPS auth blocks the CUPS vector; disabling ImageMagick delegates reduces exposure where ImageMagick calls Ghostscript.

0 / 3 Blocked