5 stories this week that change your decisions (Jun 8-14, 2026)

TL;DR Anthropic's Mythos model built proof-of-concept triggers for 13 of 14 Windows bugs Microsoft rated unlikely to be exploited, from public patches alone, and drove one to full SYSTEM control. Separately, Huawei's MPBench found that half of attacks on LLM agent memory succeed, where a fake fact planted in a document an agent reads becomes trusted memory and fires in a later session.

1. Anthropic found Microsoft's vulnerability rating system obsolete

From public patches alone, Anthropic's Mythos triggered 13 of 14 Windows bugs Microsoft rated unlikely to be exploited, and drove one to full system control. That low-exploitability rating covers 80 to 90% of even critical bugs, so the set needing urgent patching could grow about 5x.

2. Half of attacks on LLM agent memory succeed

A fake fact planted in a document an agent reads can become trusted memory and fire in a later session, no "save this to memory" command needed. Detectors built for prompt injection caught only less than half of these stealthy payloads. Protection belongs at the memory write.

3. Threat actors are using AI brands as bait in social engineering

The bait is the AI brand itself. Fake ChatGPT, Claude, and DeepSeek pages harvested credentials and card data and dropped the Vidar infostealer.

4. Anthropic wants the government to be able to block AI models. It already can.

Two days after the essay, the government forced Fable 5 and Mythos 5 offline through export controls, an early look at what such power looks like in practice.

5. Google's new audit shows 3 of 4 unlearning methods fail to forget

The only method that truly erases data keeps training on it under random labels. The clever alternatives leave fingerprints an output-only statistical test can detect. For frontier LLMs there is no affordable proof of forgetting yet: the audit itself requires a $100M+ retrain.

Sources: