Srsly Risky Biz: Chinese AI Can Learn to Love the West

This week's Risky Business newsletter covers several threads in AI and vulnerability news. Researchers at Hirundo tested Alibaba's open-weight Qwen3-Coder model and found it wrote noticeably less secure code when told it was building for a US defense contractor or a Hong Kong democracy group versus a Chinese defense contractor, echoing earlier findings from CrowdStrike and Booz Allen; Hirundo claims it can modify the model's weights to remove this pro-CCP skew, though overall code security remained poor in all cases.

Separately, last week's panic over the so-called PHP2Shell vulnerabilities in PHP proved mostly overblown initially, with early claims of mass exploitation disputed by Google and CrowdStrike, though by October 25-26 credible in-the-wild remote code execution reports emerged and the bugs landed on CISA's Known Exploited Vulnerabilities list. Around the same time, a separate critical flaw in Google's Brotli compression decoder — affecting PHP on Windows, Python, .NET, Node.js and many other platforms — was patched on October 21 and added to CISA's KEV list on October 27, with OpenAI credited for discovering it while hardening its model-training infrastructure.

The newsletter also notes that CISA's newly published four-year strategy document mentions AI only once, despite AI-driven exploit development and defense being among the most urgent cybersecurity concerns organizations currently face.