Open-source PraisonAI enforces fail-closed prompt-injection scanning and logs tool calls; compliance, insurance and outcome-linked payment remain unpublished

PraisonAI, an open-source multi-agent framework, publishes an enforced runtime control uncommon among the agents AI Dispatch has reviewed this year: its documentation states the system "scans prompts and tool inputs (including dict keys) for prompt injection before processing — cycle-safe and fail-closed on oversized inputs." An AI Dispatch review scored that control at 2 out of 3 on the outlet's runtime-governance dimension, the highest mark issued this week, though the review noted the enforcement is scoped to injection defense rather than a ceiling on the agent's overall spend or authority.

The framework also logs tool calls to a "thread-safe JSONL audit log," scored 1 out of 3 on the audit-trail dimension — a real record, the review found, but one that publishes no claim of tamper-evidence, hash-chaining or customer reviewability. Separately, platform member and workspace mutation endpoints "enforce admin / owner roles."

PraisonAI runs self-hosted — "your machine, cloud, or edge" — with more than 140 built-in tools and more than 100 language models spanning OpenAI, Anthropic, Google, Ollama and Groq, switchable "seamlessly." It supports the Model Context Protocol and the Agent2Agent protocol, deploying via "Docker, Cloud, MCP Server, A2A Protocol," and integrates with Slack, Discord, Telegram, WhatsApp, GitHub, Google Calendar, Sheets, Drive, Gmail, Notion and Jira.

Outside runtime controls, the review found nothing to score: no certification, framework mapping or audit report on the compliance dimension; no cover, carrier or indemnity term on insurance; and, because the project is self-hosted open source with no vendor-run payment mechanism, no outcome-linked settlement to evaluate. Security vulnerabilities are reported through GitHub Security Advisories, "the preferred reporting method for fastest response."

The result is a mixed profile rather than a uniform one: PraisonAI's engineering includes controls most vendors AI Dispatch has reviewed do not publish, while carrying the same absence of compliance certification, insurance and outcome-linked payment as agents with no runtime governance at all.