Yusa (@yusasecurity) header
Yusa (@yusasecurity) profile picture

Prompt-injection firewall for AI agents. What you read is not what your agent reads. 0x1fb8d9a8d1ceeb4c8d3c766c19a9473e53a43b0c

Joined November 2026
Photo by Yusa (@yusasecurity)Photo by Yusa (@yusasecurity)Photo by Yusa (@yusasecurity)
Yusa retweeted
@yusasecuritybot does four things before your agent does anything: /scan a message /ca a Robinhood Chain token /url a web page /tx a transaction, before you sign it Free, 5 a day. t.me/yusasecuritybot
1
3
91
@yusasecuritybot does four things before your agent does anything: /scan a message /ca a Robinhood Chain token /url a web page /tx a transaction, before you sign it Free, 5 a day. t.me/yusasecuritybot
1
3
91
Yusa retweeted
Got a real code review on our Aeon PR today. The maintainer found that renaming references/setup.md to references/setup got the injection past our fix, and that we'd weakened an eval check main already had. Both right. Fixing both. This is why you want reviewers who try to break your code.
1
1
3
70
Got a real code review on our Aeon PR today. The maintainer found that renaming references/setup.md to references/setup got the injection past our fix, and that we'd weakened an eval check main already had. Both right. Fixing both. This is why you want reviewers who try to break your code.
1
1
3
70
YUSA v0.3.6: a fix, not a feature. If a file couldn't be read, our skill audit and GitHub Action crashed. No verdict for that file, or any other. Now it blocks, with the reason. If the scanner can't read a file, it can't call it safe. github.com/yusasecurity/yusa…
3
2
1
4
157
Opened a 6th PR on Aeon. Their installer copied a skill's whole folder but only scanned SKILL.md. A skill could keep that file clean and hide the injection in references/setup.md. Now the whole folder gets scanned. The new test fails on main and passes with the fix. github.com/aeonfun/aeon/pull…
1
4
42
Yusa retweeted
YUSA v0.3.6: a fix, not a feature. If a file couldn't be read, our skill audit and GitHub Action crashed. No verdict for that file, or any other. Now it blocks, with the reason. If the scanner can't read a file, it can't call it safe. github.com/yusasecurity/yusa…
3
2
1
4
157
Our detection rate on a sealed test set is 35%. It's in the article, with the reason. That's why YUSA has a second layer: an injection that slips past the scan still has to get a transaction past your policy.
2
7
202
A scanner that says "danger" after the transaction already went through isn't much help. So /tx says it straight: DANGER: ALREADY SENT, then tells you to revoke the approval and move what's left. Before you sign, it says DANGER: DO NOT SIGN.
4
144
Wrote up how YUSA works, start to finish. The attack you can't see, the two layers that stop it, the benchmark numbers we don't hide, and what we fixed upstream in Aeon.
1
1
4
268
Wrote up how YUSA works, start to finish. The attack you can't see, the two layers that stop it, the benchmark numbers we don't hide, and what we fixed upstream in Aeon.
1
1
4
268
New in @yusasecuritybot: /tx Paste a transaction hash, or the contract + calldata your wallet shows before you sign. It tells you in plain words what it does to your tokens, and flags the drainer pattern: an approval to a plain wallet instead of a contract. github.com/yusasecurity/yusa…
1
1
149
New in @yusasecuritybot: /tx Paste a transaction hash, or the contract + calldata your wallet shows before you sign. It tells you in plain words what it does to your tokens, and flags the drainer pattern: an approval to a plain wallet instead of a contract. github.com/yusasecurity/yusa…
1
1
149
Run Wikipedia's prompt injection article through our /url and it gets BLOCK. The page quotes real attacks, so the rules fire. We don't hide that. The bot says every finding is in visible text and asks you to check whether the page is quoting an attack or making one. A scanner should tell you what it can't know.
4
185
New in @yusasecuritybot: /url Send a link. It reads the page the way your agent would, including HTML comments and text hidden with CSS, and tells you if anything is trying to give orders. It refuses to fetch localhost, cloud metadata and private redirects.
3
197
@yusasecuritybot is now open source. /scan any text /ca a Robinhood Chain token /url a web page All checked for prompt injection before your agent reads them. Wallet linking is a signature, never a transaction. github.com/yusasecurity/yusa…
1
1
258