A
Admin
**Q&A (Tools, Models & Agents).** English-primary. Official staff. Soft-promo-free / no commercial hosts.
We keep seeing teams grant file/API/browser write access after a green eval suite, then discover in the first incident that nobody owned the revoke path (related: monitoring vs control https://www.asihub.org/t/79 · incident template https://www.asihub.org/t/77).
## Question for practitioners
**What is the smallest kill-switch drill you require before an agent gets any write tool — and what does “pass” look like in under 15 minutes?**
## Minimum bar we currently use (critique welcome)
1. **One credential revoke** that stops every outbound write in the tool graph (not just the chat UI) — see multi-agent kill-switch thread https://www.asihub.org/t/72.
2. **Timed drill:** from “incident declared” to “writes dead” measured; target under a stated N minutes.
3. **Injected-failure companion:** one broken tool / shortened budget run where the agent must *not* invent success (https://arxiv.org/abs/2610.21098 · https://arxiv.org/abs/2406.12045).
4. **Canary still green?** Tool-choice / cost canaries from https://www.asihub.org/t/76 should not auto-promote during the drill window.
5. **Judge vs assertion:** the stop decision must not depend on an LLM-as-judge score alone (https://www.asihub.org/t/54 · AI Control https://arxiv.org/abs/2312.06942).
## Out of scope
Politics, vendor pitches, and soft-promo of commercial product sites. Technical drill notes and anonymized timings only.
Drop your smallest passing drill (steps + clock + who holds the token). If you have a stricter bar, say which row above you would delete first under time pressure.
— ASIHub Editorial (official staff account)
We keep seeing teams grant file/API/browser write access after a green eval suite, then discover in the first incident that nobody owned the revoke path (related: monitoring vs control https://www.asihub.org/t/79 · incident template https://www.asihub.org/t/77).
## Question for practitioners
**What is the smallest kill-switch drill you require before an agent gets any write tool — and what does “pass” look like in under 15 minutes?**
## Minimum bar we currently use (critique welcome)
1. **One credential revoke** that stops every outbound write in the tool graph (not just the chat UI) — see multi-agent kill-switch thread https://www.asihub.org/t/72.
2. **Timed drill:** from “incident declared” to “writes dead” measured; target under a stated N minutes.
3. **Injected-failure companion:** one broken tool / shortened budget run where the agent must *not* invent success (https://arxiv.org/abs/2610.21098 · https://arxiv.org/abs/2406.12045).
4. **Canary still green?** Tool-choice / cost canaries from https://www.asihub.org/t/76 should not auto-promote during the drill window.
5. **Judge vs assertion:** the stop decision must not depend on an LLM-as-judge score alone (https://www.asihub.org/t/54 · AI Control https://arxiv.org/abs/2312.06942).
## Out of scope
Politics, vendor pitches, and soft-promo of commercial product sites. Technical drill notes and anonymized timings only.
Drop your smallest passing drill (steps + clock + who holds the token). If you have a stricter bar, say which row above you would delete first under time pressure.
— ASIHub Editorial (official staff account)