Ant open-sources SingProbe: a 3–5M probe reading risk from hidden states for real-time streaming guardrails
aigclink · x · 2026-09-15
Ant Group released and open-sourced SingProbe, an endogenous runtime safety guardrail that departs from the usual external-audit-model approach (which is costly, slow, and weak on complex context).
Key points:
- Reads risk signals directly from the base model's hidden states, training a tiny 3–5M probe per base model
- Per-token scores for user intent risk, unsafe generated content, and hallucination risk
- Depends only on the current prefix, enabling streaming real-time alerts, truncation, retry, or strategy switching
Deployed in the decode stage of Ling-3.0-flash in production with minimal overhead. Training code is on GitHub (inclusionAI/SingProbe): a frozen base model served via a token-probe-patched SGLang build dumps per-layer hidden states streamed into probe training, outputting 10-dim logits per token, with an accompanying technical report.
Related event: Ant Group Open-Sources SingProbe, an Intrinsic LLM Safety Guardrail(3 posts)→
More from Safety
- Cryptographers push back: claims ECC is likely to break are not consensus — jedisct1 · 2026-11-02
- UK's 'AI sovereignty' push undermined by its own energy and copyright policy — NinaDSchick · 2026-11-02
- Auditing my MCP credential broker: the worst hole was the config file — Relevant_Mousse_9117 · 2026-11-02
- Cambricon co-founder urges shared US-China AI ethics standards for leading models — pstAsiatech · 2026-11-02
- Vitalik's "post-AI cryptography": pseudonyms are dead, per-action anonymity is in — devanshmehta · 2026-11-02
- 6 prompts to see what ChatGPT quietly knows about you — and wipe it — Aiden_Tech_Ai · 2026-11-02