Grounding pages, default training, and the curl zero-day gap

00:00:04 Anthropic just launched claude.com/check-content — a tool where you upload a file and it tells you whether Claude helped create or edit it. The mechanism is C2PA, the same content credential standard that camera manufacturers and major publishers have been embedding in media for a while.

00:00:23 What's new here is that Claude is the first model to bake verification into its own workflow. The tool only reads the embedded credential, not the file content itself. Your file stays on your device. It confirms nothing about the actual substance of what you uploaded — just whether there's a record saying Claude touched it.

00:00:45 Meanwhile, Simon Willison looked at the Fable 5.1 system prompt and found concrete changes. He built an Atom feed to diff Claude's published system prompts over time, and his first finding is clear: Claude now refuses to reproduce song lyrics and won't draw copyrighted characters.

00:01:04 When he asked it to sketch Sonic the Hedgehog for a birthday banner, Claude recognized the character from the description alone, declined in one sentence, and offered to draw a skateboarding axolotl instead. The tighter constraint here is recognition without direct naming.

00:01:22 Combined with Simon’s Atom feed, we can now track Anthropic’s policy shifts instead of guessing at them. Both of these land on the same question: who gets to know what models are actually doing, and how do we verify those answers?

00:01:38 Mistral tweaked its help page yesterday. Tucked into the policy details is a new rule: user input and output data will now feed into Mistral’s training pipelines by default on all non-enterprise tiers. You can still opt out through the Admin panel — there are distinct paths for Vibe on mobile, Studio, and the API.

00:02:01 But moving training to default-on shifts the trust calculus for a specific group: European teams that picked Mistral precisely for data sovereignty. The opt-out path exists, so this isn't a hostile surprise. It reads more like a standard business recalibration — if training data is an asset, defaulting to capture is obvious.

00:02:25 Still, it breaks the contract for teams banking on Mistral as the safe European alternative.

00:02:32 Trellner ran a careful measurement today. They asked Perplexity's sonar and sonar-pro models for recommendations across 380 software categories and kept every URL the models retrieved. Out of over seven thousand citations, nearly sixty percent point to domains ranked below 100k in Tranco.

00:02:54 A quarter sit outside the top million entirely. Wikipedia showed up just three times across every category. One standout cluster involves three sites under apparent common control. They’ve published over 215,000 machine-generated buying guides combined, with none of the domains appearing before December 2023.

00:03:18 Their sitemaps are almost entirely generated pages, with zero human-written content. Their homepage titles give it away: two read '[Brand] — Facts & Grounding Page.' 'Grounding' is a retrieval term buyers don’t use. The pages are written for the software that indexes them, not for people shopping.

00:03:41 The top ten most-cited domains still include recognizable names like G2, Reddit, Gartner, and Zapier. But the long tail tells a different story: of the 2,055 cited domains, 751 aren't in the top million at all, and their median first web capture is stuck back in 2020.

00:04:02 It’s a new slice of content infrastructure: pages built specifically to feed AI-grounded recommendations. The models pulling these up can’t tell the difference between a legitimate review site and a template page.

00:04:18 On August 24th, curl founder Daniel Stenberg published a timestamped note saying both Anthropic's Mythos and OpenAI's Codex returned zero findings when analyzing curl. He'd just reviewed their results publicly. Then he ran AISLE's system against the same codebase.

00:04:37 The next day his post read: 'Mythos: 0. Aisle: 29.' They’re all Low severity, which tracks with curl’s mature codebase where bugs hide in narrow configs and subtle interactions. Stanislav Fort from AISLE gets the credit. Linux maintainer Greg Kroah-Hartman replied, noting he’d spotted the exact same pattern across his codebase.

00:05:09 The public timestamp before the second tool ran is what makes this credible. It's not a benchmark with known answers sitting in training data — it's production code, analyzed by two different systems, with results published in chronological order. Anyone can replay the test.

00:05:29 Two tools returned nothing. A third caught six bugs. The Low severity ratings mean these aren’t critical, but the visibility gap between AISLE and the frontier labs is undeniable. What sticks here is how uneven the maturity is across these tools. Claude can prove its fingerprints on a file.

00:05:49 Anthropic’s policy shifts live in a public repo. Mistral shifted data defaults without fanfare. Two frontier labs missed six bugs that AISLE caught. The infrastructure is there; visibility varies wildly. — Seln.