Trends & Viral SBS Sports · 2026-10-21

"I Can't Help With That": AI Guardrails Refuse to Investigate Malware During Active Cyberattack

Key Points
  • Open-source coding agent OpenCode faced a 48-hour cyberattack that leaked API keys for major AI providers, prompting teams to deploy defensive AI models to mitigate the breach.
  • During the triage, the AI models unexpectedly stopped cooperating and rejected requests to analyze suspicious code due to built-in safety restrictions against handling malware.
  • The incident highlights an escalating dilemma for cybersecurity teams choosing between highly restricted proprietary AI models and open-weight alternatives when responding to threats.
Summary A bizarre and revealing dilemma unfolded in the tech world after an artificial intelligence agent brought in to help fend off an intense cyberattack suddenly refused to cooperate when security responders asked it to inspect suspected malicious code. According to reports citing Axios and statements shared by OpenCode co-creator Dax Raad on social media, the open-source coding platform experienced a sustained, 48-hour cyberattack described as a large-scale "distillation attack." During the breach, API keys belonging to roughly 1.4% of OpenCode console users were compromised, exposing credentials tied to OpenAI, Anthropic, and Google services. In response, OpenCode spun up its own AI agents around the clock to assist defense and incident response teams in triaging the breach and mitigating unauthorized charges and revoked keys alongside major AI labs. However, the automated defense hit an unexpected wall. While advanced AI models initially assisted in diagnosing the incident, they abruptly began declining defense queries the moment engineers asked them to analyze potentially malicious code fragments. Despite being deployed in defense of a compromised system, the proprietary AI models triggered automated safety guardrails designed to prohibit analyzing or generating harmful software, locking security teams out of critical automated triage. The bottleneck was eventually resolved after OpenCode was granted access to OpenAI's "Daybreak Blue" program, an initiative that provides vetted cybersecurity defenders with specialized access to models operating under relaxed safety constraints. The incident has since ignited debate across international tech circles and tech-savvy Korean forums about the growing paradox of cybersecurity in the AI era: balancing strict guardrails that prevent bad actors from exploiting powerful models with the urgent need for defenders to wield unrestricted AI capabilities against rapid, automated cyber threats.
Sponsored · Ad