Show HN: Bastiontrace – Forensics for prompt-injected AI agents (github.com)

Bastiontrace, a new tool in the Bastion trilogy, has been launched to enhance the forensic analysis of prompt-injected AI agents. This tool allows users to read a trace of tool calls made by an AI agent, identify instances of prompt injection, and assess the resulting "blast radius"—which includes where the injection occurred, the forbidden actions it triggered, and all related calls. Unlike other solutions, Bastiontrace operates independently without reliance on large language models or cloud services, offering a lightweight option for analyzing JSONL traces. The significance of Bastiontrace lies in its ability to provide actionable insights into AI agent vulnerabilities. By employing advanced tracing techniques, the tool assists developers in pinpointing not only the injection points but also the full chain of activities that ensued from a single injection. Key features include the ability to analyze traces for issues, generate policies to prevent future exploits, and seamlessly integrate findings into existing security frameworks. This capability is crucial for bolstering AI security in various deployments, making it easier for development teams to implement robust defenses against prompt injections and other security threats.