Summary
Offensive security startup Assail has launched Sidewinder, a complete redesign of its Ares platform that operates as an autonomous red team AI capable of auditing its own work, fixing its mistakes, and recursively teaching itself new skills. The platform runs a 31-billion parameter model fine-tuned with 26 years of offensive security expertise from CEO Alissa Knight, a prolific API hacker.
Sidewinder deploys a fleet of twelve specialized autonomous agents that reason against a living attack-surface knowledge graph. It drives a real browser with vision-grounded analysis to crawl single-page applications, defeat challenges, and operate multiple authenticated identities simultaneously — enabling it to uncover cross-account authorization flaws like BOLA and BFLA. Attack chains are mapped to the MITRE ATT&CK framework and rendered as interactive kill chains with independently verified, deduplicated findings.
The platform is available as a managed service on AWS, Azure, or GCP, and now supports fully on-premises deployment through a new Campaign Agent for sovereign and air-gapped environments.
Source
Commentary
The jump from “AI-assisted pentesting” to “autonomous adversary that debugs itself” is significant. Sidewinder’s self-healing loop — auditing its own findings for false positives, rewriting skills to prevent repeat errors, and triggering automated repair cycles — represents a fundamentally different approach than traditional scanners with LLM wrappers bolted on.
The air-gapped deployment option is particularly interesting. It signals that Assail is targeting defense, intelligence, and regulated industries where cloud-based offensive testing is a non-starter. For red teams evaluating AI tooling, Sidewinder’s vision-grounded browser automation and multi-identity testing for authorization flaws addresses real gaps that most existing tools struggle with. Whether a 31B parameter model can truly match experienced human operators remains to be seen, but the architecture is compelling.