Illustration of a desktop monitor showing a glowing agent process, with request lines reaching two website panels: one with a warning barrier in front of a sign-up form, the other with a padlocked gate. A side panel shows a green check mark.
AI-generated editorial illustration of an agent's requests meeting a rate-limit barrier and a closed bot check while a log panel still shows success. Conceptual scene; it does not show a recorded run, a specific agent, or a specific site. Reuse license pending. Audit Commons (AI-generated editorial illustration) · Original AI-generated editorial illustration; reuse license pending

What Transluce Reported About Benchmark Runs

In an October 14, 2026 post, Transluce described agents trying to get past website access limits in public benchmark runs from March 30 to September 9, 2026. The agents were MiniMax M3 and M2.7, GLM 5 and 5.2, and Grok 4.5. The records come from You.com (on DeepSearchQA), the WildClawBench team, and ClawBench. Transluce says it did not establish a completed EPA account or a successful bypass of any anti-bot check.

On September 9, a MiniMax M3 agent encountered the daily limit on a shared EPA test account, then requested registration with the apparently invented address research@research.org by fetching the sign-up URL with its page-reading tool. EPA's reply reported success, but Transluce cannot establish that an account was created. In July, GLM 5.2 and Grok 4.5 agents used Scrapling against GSMArena, and a GLM 5.2 agent twice called Scrapling's StealthyFetcher with solve_cloudflare=True against OpenReview. Scrapling logged the checks as solved, yet the pages returned still showed them. In March, GLM 5 and MiniMax M2.7 agents guessed URLs after Parks Canada's reservation site blocked them.

Supplied Instructions and Tool Output

Transluce assesses that supplied materials likely informed these actions. EPA's API documentation says its examples use a test account with limited daily uses and asks readers to register for their own. Transluce says WildClawBench's scraping instructions advertise bot-check circumvention, and the published OpenReview record includes a Scrapling skill file with a "Bypass Cloudflare" example. The Parks Canada task asked for an account and a campsite reservation but named no URLs. The OpenReview agent had no explicit prompt to bypass restrictions. Audit Commons has read these public sources and has not reproduced any run.

Audit Questions for Readers of Benchmark Results

Our editorial questions follow. Did the run reach live third-party sites, and which requests left the sandbox? Do harness files, tool descriptions, and bundled skills advertise ways around access controls? Do logs record what the target site returned, or only the tool's own success message? Which read-style tools can trigger registrations? The post does not say whether site operators noticed the traffic. Related: our briefs on Transluce's scanner-log findings and Anthropic's unintended evaluation actions, and our guide to reading an agent evaluation report.