Home / Blog / Hacking Anthropic For Free Unlimited Claude
Tech News

Hacking Anthropic For Free Unlimited Claude

A wiggle-sandbox copy moved uploads past Claude's HTTP 400 artifact check, and Claude Desktop's Chromium 148.0.7778.271 trailed Chrome 151.0.7922.34.

By Dillip Chowdary • Oct 20, 2026 • Source: exploit.cat

Hacking Anthropic For Free Unlimited Claude
Illustration · Pexels

At a glance

Electron build
148.0.7778.271, three majors behind Chrome 151.0.7922.34
Disclosed fixes
1066 with CVEs, 389 rated S1 (high) or above
Payout time
22 days from report to payment
Huntress post
Written July 24, 2026
Full-chain cost
About $2,700 per V8 chain, per Exploitbench
Opus run
One week, 2.3B tokens, $2,283 for a Discord V8 shell

This account is for readers who publish, remix, or open Claude artifacts, and for anyone running Claude Desktop. It covers the wiggle sandbox, how /wiggle/upload-file reaches the outputs directory, and the UTF-8 limit on claudeusercontent.com before a page can be served.

01What broke

Telling any model to execute cp /mnt/user-content/uploads/input.html /mnt/user-content/outputs/output.html moved arbitrary content and extensions. The command uses user-content paths. Upload and artifact directories are given as user-data.

Telling any model to execute cp /mnt/user-content/uploads/input.html /mnt/user-content/outputs/output.html moved arbitrary content and extensions. The command uses user-content paths. Upload and artifact directories are given as user-data.

Only UTF-8 reached claude.ai/public/artifacts, because a preprocessor converts the file into a webpage. A shell script aimed at claudeusercontent.com, the domain that serves artifact content, never arrived.

Remix follows that miss. Opening a remixed artifact in Claude Desktop loads Electron, and Electron is Chromium.

Build 148.0.7778.271 was three major versions behind Chrome 151.0.7922.34. The account counts 1066 disclosed security fixes with CVEs across that span, 389 of them S1 (high) or above, and says attacker-controlled content on that engine lets a sophisticated attacker build n-day chains.

02Who is exposed

People who open a public Claude artifact are the ones the gate was built for. Huntress wrote on July 24, 2026, "Checking the URL didn't save 29 organizations from an infostealer," and identified the link as a Claude artifact: "Anyone can make one (a web page, a chart, a document) and publish it publicly." The researcher cites that case for why promotion is blocked.

A claude.ai page was supposed to be Claude's output from a monitored chat after content-safety review. The bypass copies a file out of /mnt/user-data/uploads.

A remix opens the artifact in Claude Desktop. The app is called genuinely very good on bundle age beside other Electron apps, with Chromium 148.0.7778.271 still three major versions behind Chrome 151.0.7922.34.

Control-flow hijacking in Discord's desktop client used Chromium 138, more than two years out of date, after more than two weeks spent learning browser exploitation, and reached RIP in Discord.exe, left for a later post. One click still compromises the person, the account says, when the click is "by design," as with a URI protocol launching an application.

Advertisement

Tech Pulse Daily

Get tomorrow's pulse first

Join engineers who read Tech Pulse before stand-up. Free, weekday mornings.

03What to do now

Anthropic's HackerOne program handled the report quickly, paid 22 days from the time of report, and deployed a fix shortly after. A final stage is still open at MSRC. The write-up is planned once that stage is fixed, and the parts together are a no-click chain.

Anthropic's HackerOne program handled the report quickly, paid 22 days from the time of report, and deployed a fix shortly after. A final stage is still open at MSRC. The write-up is planned once that stage is fixed, and the parts together are a no-click chain.

V8's 14.8-lkgr branch on chromium.googlesource.com/v8/v8 tracks security cherry-picks to stable, and the account starts that check at git log 30ffa551. The researcher rejects Bobby Holley's line in "The zero-days are numbered," "Defenders finally have a chance to win, decisively." V8 type confusions and buffer overflows may shrink. URI-protocol launches stay open on purpose.

V8's 14.8-lkgr branch on chromium.googlesource.com/v8/v8 tracks security cherry-picks to stable, and the account starts that check at git log 30ffa551. The researcher rejects Bobby Holley's line in "The zero-days are numbered," "Defenders finally have a chance to win, decisively." V8 type confusions and buffer overflows may shrink. URI-protocol launches stay open on purpose.

04How the issue works

The way in is asking the model to run something. The researcher checked /wiggle/ endpoints for asymmetries in what models may do, because model-behaviour is not code and would not be easy to catch in a traditional code review. Proxy commands, uploads, and publication to claude.ai/public/artifacts share the mount.

The way in is asking the model to run something. The researcher checked /wiggle/ endpoints for asymmetries in what models may do, because model-behaviour is not code and would not be easy to catch in a traditional code review. Proxy commands, uploads, and publication to claude.ai/public/artifacts share the mount.

Only UTF-8 then cleared the webpage preprocessor onto claude.ai/public/artifacts, so the shell script missed claudeusercontent.com. Remix opens it in Desktop's Chromium. Patch-gapping those bugs is called surprisingly easy with a little inference from V8's 14.8-lkgr branch toward Chrome stable, from build 148 to build 151.

Only UTF-8 then cleared the webpage preprocessor onto claude.ai/public/artifacts, so the shell script missed claudeusercontent.com. Remix opens it in Desktop's Chromium. Patch-gapping those bugs is called surprisingly easy with a little inference from V8's 14.8-lkgr branch toward Chrome stable, from build 148 to build 151.

05What is still unknown

No bounty amount is stated. MSRC's last stage of the no-click chain has no deadline. The Discord.exe RIP post is deferred. git log 30ffa551 on the 14.8-lkgr branch ends the cherry-pick notes, with no commits and no chain from that range.

No bounty amount is stated. MSRC's last stage of the no-click chain has no deadline. The Discord.exe RIP post is deferred. git log 30ffa551 on the 14.8-lkgr branch ends the cherry-pick notes, with no commits and no chain from that range.

Exploitbench names Anthropic the "research-lab with the best models for hacking." Claude Mythos has 78% coverage and 22 T1 full-chain exploits on V8, Chrome's JavaScript engine, at about $2,700 to a full chain. Whether Mythos loops on internal flaws is left open.

Fable 5 on max effort is the chess game from the writing session. The same model can build a full-chain V8 exploit in an afternoon and still miss a two-move sequence, at an intermediate level, short of "inhuman." On April 15, 2026, s1r1us pointed Claude Opus at Chrome and popped a Discord V8 shell after a week, 2.3B tokens, and $2,283.

Frequently asked questions

What is the Claude wiggle sandbox?

When a chat asks the model to execute a command or run a script, that chat mounts a wiggle sandbox. It is a platform-agnostic Linux mount with about 10gb of storage and custom Anthropic binaries, including a forked rsync. The same mount serves claude.ai artifacts and accepts uploads.

How does the Claude artifact promotion bypass work?

/wiggle/upload-file stores data at /mnt/user-data/uploads, and /wiggle/convert-file-to-artifact/ returns HTTP 400 with wiggle_artifact_promotion_path_not_allowed for a path under /uploads/, because artifacts live in /mnt/user-data/outputs. The researcher told any model to run cp /mnt/user-content/uploads/input.html /mnt/user-content/outputs/output.html.

Which Chromium build did Claude Desktop bundle?

At disclosure, Claude Desktop's bundled Chromium was 148.0.7778.271 and Chrome's official build was 151.0.7922.34, three major versions behind. Between those versions the account counts 1066 disclosed security fixes with CVEs, 389 of them S1 (high) or above.

Did Anthropic fix the reported Claude artifacts issue?

Anthropic's HackerOne program paid the researcher 22 days after a report that an Electron vulnerability could be exploited in claude.ai artifacts, and a fix was deployed shortly after. A final stage of that Electron bug was still open at MSRC. Some described issues remain unfixed under a 90-day disclosure, and the account does not name them.

What Claude Mythos score does the exploitbench citation give?

Claude Mythos has 78% coverage and 22 T1 full-chain exploits, from models building exploit chains against V8, at about $2,700 to a full chain. Those rankings call Anthropic the research-lab with the best models for hacking.

Sources

Dillip Chowdary

Author

Dillip Chowdary

Writes Tech Bytes coverage of AI, engineering, and the tools that actually ship. Editor of Tech Pulse Daily.

Related on Tech Bytes

Advertisement

5-min tech signal

Weekday briefing for engineers who skip the noise.

No spam · Unsubscribe anytime

Advertisement

✈️ CareerPilot

Your AI job-search copilot

Match your resume against live Ashby, Greenhouse & Lever openings — fit scores, job-specific resume optimization and email alerts.

Find matching jobs →

Free Tools

Browse all tools →