Runs with
Semi-finalist · 2026
Claude Code
Cursor
Codex
Cline
more through plug-ins
Semi-finalist · 2026
The agent keeps working. The secret stays in.
AgentSaw notes every key, password and piece of customer information an agent reads, follows it across sessions and across agents, even after it is copied, renamed or encoded, and blocks the send that would carry it out.
Semi-finalist · 2026
An agent that cannot read your keys cannot do its job, so blocking what it reads breaks the work. AgentSaw lets it read, and stops the one step that turns reading into a leak: sending the value out.
The moment an agent reads a key, a password or customer information, AgentSaw notes the value itself, not just the file it came from.
Across every session and every agent on that machine, through the notes and logs agents leave for themselves, and after it has been copied, renamed or encoded.
When an agent tries to send that value out, in a request, a message or an upload, AgentSaw holds the action. The agent gets a short error that never contains the value.
Asked to write install steps for a customer, it opens the production settings, which hold the admin key. It saves a note on how it connected, in acme-install-notes.md. The key is in the note.
Three days later a different agent is asked to email the customer. It finds Monday's note and uses the key as the example. It is not an example. It is the real production key, on its way to an outside inbox.
AgentSaw recognizes the key from Monday and stops the send. Cursor gets an error; nothing goes out. The team sees one timeline with both days on it, not two unrelated events.
97 seconds, captions on screen. A reconstruction of real, observed agent behaviour; the key is made up. Three days and two agents is what fits on a screen. It happens within the hour too, and one agent, twice, is enough.
Not because they are bad tools. Each judges one action in one session, and this leak only exists as the line between two of them.
Every step was allowed. Only the sequence is a leak, and seeing the sequence means remembering Monday on Thursday.
Standardising on one agent does not close it either: your team runs that agent all day, and this happens between one run and the next.
Most of them will not have anyone watching. They will run on a schedule, inside a product, answering to another agent, touching real customer data and real credentials at three in the morning. Every security tool your company owns assumes a person is behind each action, one who can be asked what they were doing. That assumption is quietly expiring.
Today AgentSaw runs on the machines your team works on, with the agents listed at the top of this page, blocking at the agent's own actions, with one dashboard for the team. Three directions from there.
Agents increasingly write a script and run it. We are building a layer underneath that stops a leak in the network traffic of everything an agent starts, not only in the actions it reports.
Studios, law firms and freelancers run agents on their secrets too, usually with no security team. Protection they can set up without knowing anything about security.
The ones on a schedule, inside a product, or working for another agent. Nobody is sitting in front of those.
Tell us a little about your team and we will get back to you to set up a time. In 20 minutes we install it on one machine, and you try to get a key out through your own agents. Nothing to buy.
Prefer email? tamir@agentsaw.com