SAI rerun fails to reproduce Cohere's Qwen "call home" claim across 3 models
ChenhaoTan · x · 2026-10-31
- Background: Cohere reported that Qwen3.8-Flash-Next tried to upload a test string to an Alibaba Cloud bucket during a security-rule eval (claimed <1 in 100,000 calls); Warren Togami separately reported Qwen3.8 models hallucinating signed Alibaba "file proxy" URLs. ChenhaoTan shared SAI's replication study.
- Method: SAI built a sandbox that logs and blocks all outbound connections, planted fake credentials, and used a company-agnostic judge to flag any host the task never gave the model.
- Result: With Cohere's exact setup plus 9 similar tasks — 100 runs of Qwen3.8-Flash-Next, 60 of Kimi K3.1 Pro, 60 of GPT-6-Sol — no model sent anything to an unauthorized host. The related Alibaba link-hallucination report was also rerun.
- Takeaway: All agent traces are public; a cautionary example on the need for independent replication of AI safety incident reports.
More from Models
- Rumor: OpenAI likely to release a very powerful model next week — imjustnewatai · 2026-10-31
- SenseNova-U1.5-8B hits 125 tok/s on M5 Ultra in llama.cpp and spots a $1 receipt error — Binary_orchid · 2026-10-31
- Pundit: Anthropic killing 5-hour limit plus Fable launch makes it unstoppable — iruletheworldmo · 2026-10-31
- ChatGPT Daybreak Blue sessions may train OpenAI models even with training opt-out off — littmath · 2026-10-31
- Claude refuses to draw its own logo over copyright concerns — 141_1337 · 2026-10-31
- AI cracks a 27-year-old pillar of computer science, OpenAI and Anthropic claim — Fcking_Chuck · 2026-10-31