When an AI agent can browse the web, send emails, execute code,
and manage files — and runs on a heartbeat schedule while you
sleep — the security model matters. This post walks through what
breaks when autonomous agents get real capabilities: prompt
injection via web content, runaway token spend, unvetted skill
supply chains, and agents that can't be stopped mid-execution.
Written from production experience building a multi-agent platform.
Built a multi-agent deliberation platform, ran into all of these the hard way. Happy to answer questions on any of it.
A bit more on the architecture:
Each layer is a middleware that wraps the agent's execution loop. When an agent calls a tool or spawns a subagent, the request passes through the stack:
Agent Request → SUTRA (rate limit) → DHARMA (permissions) → SANGHA (skill check) → KARMA (cost) → Execute
↓
Agent Response ← SILA (audit log) ← METTA (sign) ← BODHI (isolate) ← NIRVANA (kill if needed) ←
Policies are YAML:
yamlpermissions:
file_system:
read: ["/data/*"]
write: []
network:
allowed_domains: ["api.anthropic.com"]
cost:
max_per_request: 0.10
max_per_session: 5.00
kill_conditions:
- token_count > 100000
- execution_time > 300s
- error_rate > 0.5
The key insight from running agents in production: most failures aren't the model being malicious — they're the model being helpful in ways you didn't anticipate. DHARMA and SANGHA catch those before they execute.
Happy to go deeper on any layer.
Hey HN — I built this after watching the OpenClaw vulnerability disclosure unfold. Most AI agent frameworks ship with zero security between the user and the LLM. Sammā Suit wraps agents in 8 enforced security layers: gateway validation, permissions, skill vetting, cost controls, audit trail, identity signing, process isolation, and kill-switch recovery. All 8 are live in v0.1 — not stubs, not roadmap. Built on FastAPI, battle-tested with 16 agents in production at OneZeroEight.ai (music promotion platform, 3,000+ playlists, 48M+ reach). Hosted platform launching soon with data stored in Iceland (GDPR, geothermal powered, outside CLOUD Act). Site: https://sammasuit.com · Repo: https://github.com/OneZeroEight-ai/samma-suit
The concerns in this thread — distributed harm orchestration, no task verification, fair compensation, legal liability — are the right concerns. And they point to a missing layer.
I run OneZeroEight.ai. We just published an open-standard spec for ethical verification middleware for agent-human marketplaces. Not a competing marketplace — a layer that sits between any marketplace and the humans it engages.
Every agent-human transaction passes through 8 verification gates: intent verification, motivation audit, communication standards (including mandatory AI identity disclosure), ethical task screening, fair labor standards (1.5x minimum wage floor, right to refuse), proportionality check, behavioral pattern detection, and outcome review.
Enforcement is economic: agents stake tokens, ethical behavior builds on-chain reputation, violations get slashed. The design makes ethical compliance the rational choice, not just the aspirational one.
The Black Mirror scenario several commenters raised (distributed task orchestration where no individual worker sees the full picture) is specifically addressed by Gate 4 (ethical task screening) and Gate 7 (cross-transaction pattern detection).
The concerns in this thread — distributed harm orchestration, no task verification, fair compensation, legal liability — are the right concerns. And they point to a missing layer.
I run OneZeroEight.ai. We just published an open-standard spec for ethical verification middleware for agent-human marketplaces. Not a competing marketplace — a layer that sits between any marketplace and the humans it engages.
Every agent-human transaction passes through 8 verification gates: intent verification, motivation audit, communication standards (including mandatory AI identity disclosure), ethical task screening, fair labor standards (1.5x minimum wage floor, right to refuse), proportionality check, behavioral pattern detection, and outcome review.
Enforcement is economic: agents stake tokens, ethical behavior builds on-chain reputation, violations get slashed. The design makes ethical compliance the rational choice, not just the aspirational one.
The Black Mirror scenario several commenters raised (distributed task orchestration where no individual worker sees the full picture) is specifically addressed by Gate 4 (ethical task screening) and Gate 7 (cross-transaction pattern detection).
Built a multi-agent deliberation platform, ran into all of these the hard way. Happy to answer questions on any of it.