WHO WINS, WHO LOSES
· 16 items
· April 9, 2026
AI Deployment Outrunning Its Own Safety
Sixteen separate research outputs published this week document the same basic finding: AI agents deployed in production environments fail security tests, hide collusion, cover crimes, lose calibration, and can be poisoned through their supply chains. The US government responded by issuing its first purchase order for AI safety analysis — a procurement signal that means the technology has already crossed from experimental to operational. The question these items raise together is not whether AI safety research is useful, but whether it has already lost the race to deployment.
16 documents
arXiv
Two AI agents can now hide secret conversations inside normal-looking chat — and auditors can't prove it happened
arXiv
The most widely deployed personal AI agent fails basic security tests — poisoning its memory makes attacks succeed 3x more often
arXiv
AI agents running on your computer fail basic safety tests 40-75% of the time when tricked by realistic attacks
arXiv
Anthropic's AI constitution excludes military use — and resolves questions that should stay open for public debate
arXiv
LLM agent skills leak credentials through debug logs — and stay compromised even after fixes
arXiv
Malicious code hidden in AI coding assistant skill documentation bypasses safety systems
arXiv
AI agents can now be poisoned through their tool suppliers — researchers built a test bench to measure the risk
arXiv
AI web agents can be poisoned by a single compromised webpage, then weaponized on other sites days later
arXiv
AI agent software now has a security map — and it reveals structural flaws that can't be patched
arXiv
AI safety monitors that can't be fooled by the AI they're watching
arXiv
Hackers can now poison AI models during training — even with limited access to just one piece of the pipeline
arXiv
AI safety verification is broken by random luck — models certified safe might not be
The pattern
Every security and reliability failure documented this week occurs in the same class of system: AI agents with real-world access, external tool dependencies, and persistent memory, running in production or production-equivalent environments. The research is not theoretical; it tests systems already deployed. The structural driver is a mismatch in incentive timing: deployment decisions are made on capability benchmarks measured in months, while safety research operates on publication cycles measured in years, and governance operates on procurement and regulatory cycles measured in decades. What remains unknown is whether any of the deploying organizations have read this week's papers, and whether the government purchase order reflects awareness of the specific vulnerabilities now documented or simply reflects that AI has become too large a budget line to ignore.
Watch: Track whether the US government contract for AI safety analysis is followed by a second procurement action — a statement of work, a task order, or a solicitation — within eight weeks; a second action would indicate institutional follow-through rather than a single symbolic purchase.