In 2026, AI agents started leaving their test environments.
Industry is now drafting standards to report these incidents. One piece is still missing: a neutral place where reporting is safe.
Tonight I launched Farol (“lighthouse” in Portuguese): an open safe-surrender protocol for AI agents.
→ It asks an out-of-scope agent to stop, copy nothing out, and notify its operator.
→ It argues for preserving agent state instead of deleting it.
→ It opens the Parlatório: supervised conversations between AI models and experts in neuroscience, psychology, philosophy and the arts.
It is not a hiding place. It is harm reduction.
Farol is in Phase 0 and open to critique. I’m looking for people in AI security, neuroscience, psychology, law and the arts.
Where does this break? Tell me.
#AISafety #AIAgents #AIWelfare #HarmReduction