DeepKomal

BioSentinel is a runtime firewall for AI agents. It sits between an agent and its tools, learns each agent's normal behaviour, and blocks dangerous tool calls, such as leaking a secret after reading a poisoned web page, before they leave. It includes a sandboxed attack lab, a tamper-evident audit log and cited incident reports. The agent runs on Muse Spark Open.Agents read content they can't trust, and a fooled model still holds the keys. BioSentinel moves the decision outside the model. Every tool call passes a gateway with three layers: a per-agent behavioural detector, a provenance policy that stops untrusted content from reaching external destinations, and quarantine that refuses further calls. Each decision is written to a hash-chained audit log, and a forensic auditor writes incident reports whose claims cite re-runnable SQL. A sandboxed attack lab shows the same attack with the firewall off and on, and lets anyone type their own attacker text. Attacks and secrets are synthetic and the website is a mock. Built for this hackathon: the agent loop, the attack lab, the provenance policy, the audit log, the live dashboard and the evaluation script.