We find what others miss, and ship the fix before the CVE is filed.
Research, open-source disclosures, and upstream fixes — the full record of everything we've shipped. Back to the overview →
Catching known attacks is table stakes. The real signal is finding attack classes nobody has named yet: this URL-smuggling trick is the first of a research pipeline, not a one-off.
Real bugs, fixed in the open.
We found a brand-new way to smuggle a banned web address past a security check — no name, no public record until we named it. One discovery, six public moves: free rules released, fixes merged into the scanners developers run, and the report sent upstream first.
Hover over (or tap) the i on any card for a plain-English explanation of what it is and why it matters.
Security checks usually keep a blocklist of banned web addresses. We found a way to write a banned address in a disguised form the check doesn't recognize, so it slips straight through. This is the original discovery the other five cards all build on; here, we wrote the rules that spot the disguise and published them free for anyone to use.
When we found it, this attack had no name and no public record. Most security companies would keep a discovery like that private and sell access to it. We gave the fix away instead, so any team can catch the attack whether or not they ever pay us.
Open-sourced detection library
We found a new way to sneak a banned IP address past a security check, then open-sourced the rules to catch it. Free for anyone to use.
CodeQL is the security scanner built into GitHub; it reads through a project's code looking for known weaknesses before they ship. We wrote a new rule for it that recognizes the address-disguising trick.
Once the rule is merged, every project that runs GitHub's scanner gets the check for free, automatically. The change is public, so you can read the exact rule we wrote yourself.
Detection rule shipped to CodeQL
We added a rule to GitHub's own code-scanner so it flags this attack for every project that runs it. The change is public.
Not everyone uses GitHub's scanner. Semgrep is a second popular one, run by a lot of teams who use other setups. We submitted the same catch to it as well, as its own rule.
Shipping the same fix to both of the big scanners means far more projects are protected by default, instead of only the ones using one particular tool.
Detection rule shipped to Semgrep
We gave the same catch to Semgrep, the scanner millions of developers run outside GitHub, so it becomes a default there too.
A2A is a new standard that lets AI agents hand work to each other. While stress-testing it, we noticed the official toolkit kept background jobs running even after it was told to shut down, a quiet leak. We wrote the fix and sent it to the people who maintain it.
A tool that doesn't stop cleanly can leak data or pile up wasted work in the background, where nobody is looking. Fixing it in the official toolkit protects everyone who builds on it, not just us.
Bug fix shipped to the A2A agent SDK
While stress-testing agent-to-agent protocols we found the official A2A toolkit leaking background tasks on shutdown, and sent the fix upstream ourselves.
We wrote a tool that automatically rewrites the vulnerable code to add the missing safety check, and contributed it to Uber's open-source gopatch project.
Instead of asking thousands of developers to hand-fix the same flaw, this fixes it for them automatically, everywhere it appears.
Auto-fixer contributed to gopatch
We wrote a tool that automatically rewrites vulnerable Go code to add the missing check, and sent it to Uber's gopatch project.
Before telling the public anything, we reported the original address-disguising flaw privately to the team that maintains the Go programming language, and gave them time to respond.
Giving the maintainers the first chance to fix it, before attackers learn the details, is what responsible disclosure means. The Go team decided guarding against it was the app developer's job, so we published the free detection rules ourselves to close the gap.
Reported upstream to the Go team
We told Go's security team first; they called it the caller's job to guard, so we shipped the public rules ourselves.
Get this depth on your own pipeline.
The same research that finds these attack classes ships in the Gate. Request access, or re-verify a signed verdict yourself first.