LLM distillation demand turns API keys into prime targets for attackers.

The incident wave. The incidents came in two flavors on October 21. Node.js, GitLab, and Oracle shipped security fixes, while a critical PHP vulnerability (CVE-2025-?) with a 9.2 CVSS score was disclosed. Simultaneously, OpenCode and Railway were attacked, with API keys stolen. The speaker stresses that CVEs are about patching known weaknesses, while attacks are active exploitation, and each demands a different response.

The PHP situation. The PHP situation is the most alarming. The FBI issued a public advisory, urging organizations to take systems offline until patched, warning of remote code execution that can give root access. The vulnerability allows an attacker to upload a webshell and control the server. The embargo had already broken: Chain Guard released patched container images hours early, letting attackers study the fix ahead of widespread deployment. This is a textbook example of how disclosure timing matters.

OpenCode and Railway. OpenCode and Railway both had environment variables compromised. OpenCode's API keys were consumed within minutes, with some users seeing charges up to $3,000 and OpenAI accounts suspended. Railway's breach included encrypted environment variables, but the platform also took down deployments for 16 hours, leaving the speaker unable to redeploy with rotated keys. The speaker's frustration with Railway's operational response is a second lesson: even after a breach, recovery can be hampered by the provider.

The real driver. The real driver, the speaker argues, is less about traditional cybercrime and more about model distillation. Anthropic and OpenAI have banned competitors, especially Chinese labs, from scraping their output via API. To train models like DeepSeek, these labs need massive volumes of Claude and GPT responses. They can't buy tokens legitimately, so they buy stolen keys.

The panic over PHP. The panic over PHP's '70% of websites' statistic is misplaced. Most of those are abandoned sites with negligible traffic. The real danger lies in internet-facing systems that are actively used and often run outdated PHP. The speaker's analogy: judging that PHP is everywhere by URL count is like counting every human who ever lived and averaging screen time. Leaders should inventory where PHP actually gets traffic, not just where it's installed, and focus patching there.

The core shift. The core shift: AI models can now analyze patches and generate exploits within hours, shrinking the vulnerable window dramatically. Teams need automated dependency scanning, strict secret rotation, and monitoring for abnormal API usage. The speaker's 'start of the end' framing is hyperbolic, but the direction is real: the security economy now includes a black market for inference, and any platform with keys is collateral damage. The takeaway is unambiguous: rotate keys, patch fast, and assume your secrets are never safe.