Advisory

PYSEC-2026-3507

Details, part 2 of 2. The advisory

Details

  • the published patched version v4.5.128 still reproduces;
  • latest release v4.6.57 still reproduces;
  • current main still reproduces;
  • the PoV proves unauthorized agent invocation and side effects, not only instruction disclosure.

This is distinct from private PRAI-CAND-003 / GHSA-x8cv-xmq7-p8xp, which
covers praisonaiagents.AgentTeam.launch() routes. This report covers
praisonai.app.AgentOS and AgentApp alias routes.

Impact

If an operator exposes an AgentOS app on a reachable interface, any client that
can reach it can:

  • enumerate deployed agents through GET /api/agents;
  • read agent names, roles, and instruction snippets;
  • invoke the default agent or a named agent through POST /api/chat;
  • trigger downstream tools, private context reads, memory accesses, API integrations, browser actions, or other side effects attached to the agent;
  • consume model/API budget through repeated invocation.

The exact downstream impact depends on the deployed agents. The framework-level
boundary failure is that a production deployment surface exposes agent control
without authentication and defaults to binding on all interfaces.

Suggested Fix

Use the same security model already applied to generated API deployments:

  • add authentication fields to AgentOSConfig / AgentAppConfig;
  • default auth to enabled;
  • default bind host to 127.0.0.1;
  • reject no-auth and wrong-bearer requests for GET /api/agents and POST /api/chat;
  • fail closed for non-loopback binds unless auth is configured or an explicit unsafe development opt-out is set;
  • avoid returning instruction text from unauthenticated metadata endpoints;
  • add regression tests for no auth, wrong bearer, correct bearer, and external bind without auth.

Maintainers can either update GHSA-pm96-6xpr-978x with the corrected affected
range and broader impact or publish a separate incomplete-fix advisory.

Suggested Severity

Suggested severity: Critical.

The Critical score matches the unauthenticated agent-control model: network
attacker, low complexity, no privileges, no user interaction, and high
deployment-dependent impact when agents are connected to tools, private data,
or cost-bearing services. If maintainers score only a minimal no-tool demo
agent, the impact may be lower, but the current default framework behavior is
still unauthenticated agent invocation.

Sources & licences