AI attacks demand a mental shift (softbeehive.com)

On November 13, 2025 Anthropic said a Chinese state-sponsored group used Claude Code to run an autonomous espionage campaign — a claim that drew immediate skepticism. Security researchers and outlets like Bleeping Computer criticized Anthropic’s post as vague and lacking indicators of compromise, with some calling it a PR stunt. That backlash obscured the important, less sensational point: whether or not this specific report proves a breach, the real risk comes from how LLMs are being applied, not the models alone. The technical danger is orchestration — models acting as context-aware coordinators that chain tools, automate workflows, and remove friction so attacks scale and run faster than human operators. The Darcula phishing analysis shows how a relatively simple innovation (an automated site generator with an admin panel) dramatically boosted criminal efficiency; AI can be the next such multiplier. Defenders should shift from policing models toward hardening ecosystems: invest in usability and security of open-source tooling, better detection and telemetry, and funding for practical defensive tools. Short-term controls (logging, censorship) help, but long-term resilience requires tackling automation, reducing attacker convenience, and closing the massive knowledge and tooling gaps in incident response.