This release includes 1 security fix for security teams reviewing exposed deployments.
Published 26d
MCP Browser & Automation
✓ No known CVEs patched
This release patches 1 known CVE
Topics
ai-agent
ai-agents
browser-automation
claude
docker
fastapi
+6 more
llm
local-first
mcp
novnc
playwright
self-hosted
Affected surfaces
deps
Summary
AI summaryDependency upgrades resolve all open Dependabot alerts.
Full changelog
Added
- Operator dashboard: run replay view. Open a completed agent run by job id to see its action order, approvals, final status, and screenshot artifacts, reusing the existing
/agent/jobsand/approvalsendpoints. All untrusted run data renders via text nodes and safe cell helpers (neverinnerHTML). - Operator dashboard: auth profile setup wizard. A four-step flow — name a profile and start a login session, complete login by hand in the takeover window, save the captured auth state as a named profile, and reopen a session from any saved profile.
- Local fixture server + optional live execution (
scripts/fixture_server.py,scripts/fixture_live.py): serveevals/fixtures/over loopback and drive the real controller (create session → navigate → observe) against a fixture. Opt-in; requires Playwright browsers and never runs in default CI. - WebArena Stage 0 executable contracts (
benchmarks/webarena/): typedTaskContracts parsed from the manifest, an environment-revision pin (null until a reviewed SHA is set), and a runner withvalidate/executemodes that materializes the trace/actions/screenshots/model_decisions evidence layout. Lane stays tracked-only until pinned. - Verifier lane adapter (
benchmarks/adapters/verifier_adapter.py): maps anAgentRunResultinto the CUAVerifier and Online-Mind2Web evidence lanes. Never scores —verifier_resultis alwaysNoneand records arescored: false. - Closed-tab recovery fixture + regression (
closed-tab-recovery): the fixture-eval mandatory set and a controller test now cover closing the active tab and recovering to a usable, foregrounded tab. - MCP resources & subscription examples (
examples/mcp-resources.md) with a doc-sync test that keeps the documented URIs, methods, and error codes aligned withmcp_transport.py. - Live-free coverage for the provider base layer (readiness checks, decision-parse ladder, error extraction), the stealth timing/fingerprint layer, and the mesh nonce replay cache. Controller coverage ratcheted upward while keeping the 80% gate green and avoiding live browser/network dependencies.
- Scheduled dependency audit is provided by the existing
dependency-auditCI job (pip-audit) plus GitHub Dependabot alerts.
Changed
- Startup warns when
API_BEARER_TOKENis unset in non-production. Production already hard-fails; non-production now surfaces a runtime-policy warning so a reachable dev/staging instance is not silently served unauthenticated. - Provider decision-parse ladder narrows its fall-through handlers from bare
except Exceptionto(ValidationError, ValueError), so a genuine bug propagates instead of being masked as a parse miss.
Fixed
- Codespaces stack no longer double-binds ports.
docker-compose.codespaces.ymlnow tags itsportslists with!overrideso they replace the base127.0.0.1bindings instead of appending0.0.0.0on top (which failed with "address already in use" and broke the devcontainerpostStartCommand). GET /sessionsno longer 500s after the browser is closed from VNC (external fix, thanks @gmother): session summary detects a disconnected page and marks the sessioninterrupted/live:falseinstead of raising.
Security
- Resolved all open Dependabot alerts by bumping pinned dependencies:
starlette1.0.1 → 1.3.1,cryptography46.0.7 → 49.0.0, plusredis8.0.1,pyotp2.10.0,prometheus-client0.25.0, and Playwright/GitHub Actions updates.
Security Fixes
- Dependency bumps resolve open Dependabot alerts: starlette 1.0.1 → 1.3.1, cryptography 46.0.7 → 49.0.0, redis 8.0.1, pyotp 2.10.0, prometheus-client 0.25.0
Weekly OSS security release digest.
The CVE patches and breaking changes that affected production tools this week. One email, every Sunday.
No spam, unsubscribe anytime.
Share this release
About LvcidPsyche/auto-browser
Give your AI agent a real browser — with a human in the loop. Open-source MCP-native browser agent.
Related context
Beta — feedback welcome: [email protected]