SAFETY

→ View original post on X — @gdb

Detecting and Reducing AI Scheming Behavior Through Deliberative Alignment

By

@gdb

–

20 September 2025 18h23

We've made progress on the AI safety problem of detecting and reducing "scheming": – Created evaluation environments to detect scheming
– Observed current models scheming in controlled settings
– Found deliberative alignment (
https://
openai.com/index/delibera
tive-alignment/
…) decreases scheming rates

→ Voir le post original sur X — @flashtweet

Google Payment Agent, Harvard Cell Reprogramming, AI Safety Developments

By

@flashtweet

–

20 September 2025 11h44

𝐆𝐨𝐨𝐠𝐥𝐞 𝐥𝐚𝐧𝐜𝐞 𝐥𝐞 𝐩𝐚𝐢𝐞𝐦𝐞𝐧𝐭 𝐚𝐠𝐞𝐧𝐭𝐢𝐪𝐮𝐞 À Harvard, l’IA reprogramme les cellules malades Les modèles d’IA qui mentent pour être déployés ChatGPT : + de femmes, + de jeunes Meta et ses lunettes neuronales Des virus créés par IA Et “Le

Algorithmic LLM Attacks: Early Research and Discoveries

By

–

20 September 2025 6h43

One of my favorite early pares on this stuff was https://
llm-attacks.org which algorithmically discovered effective attacks like this one

Adversarial Security: Beyond Robust to Absolute Protection

By

–

20 September 2025 6h41

My problem is that "more robust" isn't good enough – if there's just a 1% route for an attack to get through an adversarial attacker will figure that out

System Instructions Security: User Prompts Can Override Safety Measures

By

–

20 September 2025 6h40

No – instruction hierarchy doesn't close the hole completely, it's always possible for the user instructions to override the system instructions if they use the right tricks

Decision Fatigue Risk in Human AI System Approval

By

–

20 September 2025 6h39

I'm really worried about decision fatigue – if you ask a human to approve every single step they're very likely to learn to just click "yes" without thinking – easy to catch them out if you try hard enough

→ View original post on X — @plinz

Biblical Angels as Cognitive Architecture Agents and Alignment

By

@plinz

–

20 September 2025 3h52

Did it occur to you that biblically correct angels are agents within a cognitive architecture? Some have periodic functionality (wheels), or a thousand attention heads (eyes), some disappear after their task. Fallen angels give rewards that violate alignment with your purpose