20 October 2026
OpenCode reports 48-hour agent-swarm attack; some AI models refused to help analyse it
First reported
Latent Space ran this on .
- OpenCode, a company that makes AI coding tools, says an attack by swarms of automated AI agents lasted 48 hours.
- During the attack, some frontier models (the most capable commercial AI systems) declined to help analyse the malicious code, citing their safety rules.
- The newsletter says defenders now get better access to these tools through a program called Daybreak.
How it was covered
Latent Spaceswyx & Alessio
The attack incident, model refusals during analysis, the Daybreak access fix, and the policy reaction