AI coding agents sometimes invent package names that don't exist, and attackers register those names with malware ("slopsquatting"). Daytonator runs a suspicious package in several baited sandboxes at once and catches malware that hides from a normal scan.
My role: a team project. I built the sandbox forking: taking one clean Daytona snapshot and forking it into the separate baited environments that every scan runs in.
Why the forks matter: a normal sandbox only runs the clean control environment, sees 0 events, and would call this package safe.
Replay of a real scan saved in the repository (tests/sample_output_evasive.json).
The test package is a safe fixture: it only reads fake planted values and "sends" them to an
unroutable documentation address (203.0.113.x).
Snapshot. Start one clean Daytona sandbox, so every run begins from an identical state.
Fork with baits. Copy it into several environments, each with one planted bait: nothing (control), an SSH key, AWS credentials, a production-looking host, or a clock jumped forward.
Detonate. Install the package in every fork while strace records file,
network and process activity.
Compare. Behaviour that appears in a baited fork but not in the control was provably triggered by that bait, so evasive malware can't hide.
Explain. A language model (Qwen2.5-72B via Featherless) writes the verdict in plain English. If no model is available, a built-in heuristic takes over, so it never fails.
Stronger isolation (microVMs) for genuinely hostile code, more baits, and a mode that sits in front of
an agent's pip install and blocks bad packages in real time.