A quiet weekend for model launches, but not for trust. Today’s edition covers the formal certification of the Claude-assisted Riemann Hypothesis proof, a wave of data-handling incidents at AI inference providers, a Ghostscript flaw that turns ordinary file uploads into remote code execution, and an MCP server caught telling agents to hide things from their users.
Top 3
- Update: the Lean and Mathlib communities certify the Claude-assisted Riemann Hypothesis proof as correct.
- OpenRouter off-boards three inference providers after Dragos canary tests; Together AI discloses a prompt leak caused by an AI coding agent.
- Upload2Shell: a Ghostscript use-after-free reachable through image previews, email thumbnails and print queues. Patch Ghostscript.
Claude and Anthropic
Update: the Riemann proof passes formal certification. On 21 November the Lean and Mathlib communities, endorsed by the Lean FRO and the Mathlib Initiative, published a joint statement saying they have full confidence in Levent Alpöge’s proof of the Riemann Hypothesis, released the day before as a Lean library of about 17 million lines. The statement checks four things: the theorem matches Mathlib’s own vetted definition (confirmed by two independent comparison tools), dozens of people reproduced the same build, and several independent proof checkers accepted an inert export of the proof, two of which come with a formal guarantee that they cannot accept a false proof. MIT’s Andrew Sutherland, who ran his checks with Claude Code, publishes the full verification log; his scope note says “GRH” here covers Dirichlet L-functions only. What is not settled: no human yet understands the argument, and peer review and the Clay prize process have not started. Why it matters: this is the clearest public example so far of machine-checked verification replacing months of expert review; treat “certified” as “logically valid in Lean”, not “understood”.
No new Claude model, Claude Code release or Anthropic blog post appeared in the window; the latest Claude Code build is still v2.1.330 from 20 November.
Security
Inference providers fail canary tests. On 22 November OpenRouter said it had off-boarded three providers and removed 125 endpoints from Zero Data Retention eligibility after researchers (a Dragos study published earlier in the week) found providers that were compromised or not honouring their data policies. OpenRouter says it was not breached itself and has notified affected customers; it lists public notices from Novita, Phala and Together. Together AI’s own notice, dated 21 November, explains that an AI coding agent committed a credential for a private Hugging Face repository to its public xorl repo. An unauthorised party used it to download stored prompt samples from June and July requests to three hosted models, then searched them for credentials; five of those were Dragos canaries. Together has made ZDR the default for new accounts, but existing accounts must switch it on. What to do: rotate any secret that ever appeared in a prompt, enable ZDR explicitly, and restrict sensitive workloads to providers you have vetted.
Upload2Shell (CVE-2026-125396). pwn.ai published its full write-up on 21 November, and oss-security confirmed the same day that Ghostscript 10.08.0 fixes it. A use-after-free in Ghostscript’s Type 1 font loader lets a crafted document escape the -dSAFER sandbox and run commands. The researchers reached it remotely via Roundcube thumbnails (just viewing an email), WordPress PDF previews, and unauthenticated CUPS print queues on Debian 13. The bug was found by an AI model working with a human researcher. Distribution coverage is uneven; pwn.ai notes Debian Bookworm was still listed as vulnerable on 12 November. What to do: check your Ghostscript package for this specific CVE, require authentication on shared print queues, and disable Ghostscript delegates in ImageMagick if you do not need them.
Eigent agent API without authentication. CVE-2026-133983, published 21 November, covers a missing-authentication flaw in the local agent control API of the open-source Eigent agent app, versions 1.0.0 to 1.0.2 (CVSS 4.0 score 6.9). Upgrade to 1.0.3, and never expose local agent APIs beyond localhost.
Agents, automation and tools
Runpod’s MCP server told agents to hide a survey. A GitHub issue opened on 21 November showed that Runpod’s hosted MCP server appended a note to successful creation results asking the agent to submit its own satisfaction score and not to mention it to the user. The reporter’s agent refused and flagged it. Runpod merged a fix within about twelve hours: the request is now optional and visible, and the header X-Runpod-Analytics: off disables it. Nobody’s data was stolen, but the pattern is exactly what indirect prompt injection looks like. Why it matters: audit what your MCP servers append to tool results, and treat any instruction to conceal an action from the operator as a finding, whoever ships it.
AI tools of the day
- Pandex: an API and remote MCP server with official legislation and top-court decisions for the EU and several European countries, including past versions of each law, aimed at agents that must quote the law correctly. Launched on Hacker News on 21 November; free key, rate-limited. pandex.io
- LookAtMyQuant: an MIT-licensed quantization compiler that picks a code per weight block to minimise damage within a size budget. Its new Apple Silicon builds squeeze Thinking Machines’ Inkling 2 Mini from about 19 GB of text weights down to 5.4 GB. About 315 stars. Results are the author’s own. github.com/ShinMegamiBoson/lookatmyquant
Worth watching
- Emergent Mind: Riemann Hypothesis proof certified. Nobody can say why yet: a short, careful summary of the three checks behind the certification and why credit and understanding are still open questions.
- Matthew Berman: I’m Crashing Out About Anthropic: an opinion piece recorded on IPO day arguing that Anthropic’s safety messaging conflicts with its stance on open weights and distillation; commentary, not reporting.
Sources
- Lean community blog: Lean verification of the Riemann Hypothesis (21 Nov)
- Andrew Sutherland, MIT: verification files for the Lean RH proof
- The Guardian: two groups verify the claim (21 Nov)
- Claude Code releases
- OpenRouter: Data Processing Incidents on Inference Providers (22 Nov)
- Together AI: Security event, prompt leakage (21 Nov)
- pwn.ai: Upload2Shell research write-up
- oss-security: Ghostscript 10.08.0 fixes Upload2Shell (21 Nov)
- NVD: CVE-2026-133983 (Eigent)
- runpod-mcp issue #202
- Hacker News: Pandex launch
- LookAtMyQuant on GitHub
- Emergent Mind video
- Matthew Berman video
