<?xml version="1.0" encoding="UTF-8"?><oembed><type>video</type><version>1.0</version><html>&lt;iframe src=&quot;https://www.loom.com/embed/7862d9947fbc4135ac299d23f08cb37e&quot; frameborder=&quot;0&quot; width=&quot;3340&quot; height=&quot;2505&quot; webkitallowfullscreen mozallowfullscreen allowfullscreen&gt;&lt;/iframe&gt;</html><height>2505</height><width>3340</width><provider_name>Loom</provider_name><provider_url>https://www.loom.com</provider_url><thumbnail_height>2505</thumbnail_height><thumbnail_width>3340</thumbnail_width><thumbnail_url>https://cdn.loom.com/sessions/thumbnails/7862d9947fbc4135ac299d23f08cb37e-d528c472f977e6f0.gif</thumbnail_url><duration>180.114</duration><title>Don&apos;t Take Your Coding Agent&apos;s Word for It — Meta Hackathon Demo</title><description>This Loom explains Bitness, a recorder and verifier that detects spoofed tool calls and serious breaches in coding agents. It describes a simulated attack against Meta’s Muse code agent using a decoy token from a fake service, where Bitness watches externally, saves the token to disk, and stops the run based on a stop rule. It also covers Muse Spark 1.3 and Muse Glimmer 30B recording and replaying real calls with deterministic rechecks, producing plain-English advisory outputs. Finally, it notes verification can fail in milliseconds when a signed receipt field is altered, and that the system uses multiple local and MCP read-only tools to enforce signed verdicts.</description></oembed>