BOSSAI: Inside the AI attack mesh that targeted Taiwan’s emergency services

Acronis has found an “AI attack mesh” called BOSSAI, which is the sort of phrase you get when someone bolts “AI” onto a knitting circle and points it at Taiwan’s emergency services. The NCSC’s 2026 annual review we noted recently reached for “highly likely” about AI boosting cyber crime; Acronis has now supplied the show-and-tell. The mesh is built from open-source and “abliterated” models alongside commercial and closed-source systems, each handed specialised offensive functions, all held together by MCP servers. It is not a lab leak; it is a supply chain with a GitHub account.

The star turn is CyberStrike, described on its GitHub page as “the first open-source AI agent built for offensive security” — a milestone that belongs on a plaque in a museum of poor decisions. According to Acronis, it appears to have run vulnerability assessments and exploitation attempts against live targets, though the specific outputs could not be independently verified. So the sales pitch is: highly automated sophisticated attacks, provenance optional, peer review pending. “Abliterated” is the word of the week, by the way — models with the safety training sanded off, now available in a convenient mesh near you.

The genuinely funny bit is the phrase “highly automated sophisticated attacks”, delivered as if automation were the scary part rather than the fact that someone pointed it at emergency services. Acronis cannot verify the outputs, which is reassuring in the way a fire alarm with no batteries is reassuring. The open-source agent, the commercial models, the MCP servers: a modular kit for people who find “responsible disclosure” a bit much. The NCSC said AI would lower the bar for cyber criminals; BOSSAI suggests someone has also removed the bar, the door and the wall.