v0.6.0

unplug-ai 0.6.0

Minor release since 0.5.2: adversarial security fixes, public limits/judge API, and a durable local test harness.

Security

• Safe-prefix cache boundary overlap + source/policy-scoped keys (#87)
• Per-request scanner allowlist no longer sticks across scans (#88)
• Guard config mode and strict allowlist loading (#89)
• Judge action/score enforcement (#90)
• Review-debt hardening for cache, judge, ML lock, and checkpoints (#91)

Added

• Public LimitConfig / BYOLLM judge API surface (#80)
• Injection regex expansion + bidi normalizer stripping (#80)
• Local harness: test-frameworks, test-ml-harness, smoke-ml-hooks, test-all-local + TESTING_HARNESS.md (#92)

Fixed

• Post-merge review follow-ups from promote/Phase C (#81)

See CHANGELOG.md for full details.

v0.5.2

unplug-ai 0.5.2

Patch release: robustness, usability, and YAGNI cleanup after 0.5.1. Python 3.13 [ml]/[all] install was already fixed in 0.5.1.

Fixed

• Model store / Guard ML: Corrupt manifests no longer crash Guard or unplug-models; checkpoint validation, atomic downloads, clearer CLI errors (#72).
• ML inference: Safe BIOES decode, validated label maps/devices, thread-safe model load (#73).
• Agent usability: Accurate Guard.init() docs; unplug.api.* guidance for server/MCP; REVIEW outcomes log at INFO (#74).

Changed

• Packaging extras / install docs polish; wider transformers constraint; unknown active_model raises ConfigError (#75).

Removed

• Dead scrape/content/haystack shims and no-op judge config fields (#77).

Added

• Deprecation shim coverage; synthetic BIOES test fixture; CI wheel-only [ml] resolve on 3.13 (#76).

See CHANGELOG.md for full details.

v0.5.1

unplug-ai 0.5.1

Patch release: Python 3.13 packaging fix and supply-chain hardening of the bundled model catalog.

Fixed

• Python 3.13: unplug-ai[ml] / [all] now install without a Rust toolchain — transformers constraint widened to >=4.44,<5.13 so tokenizers resolves to a cp313 wheel (0.5.0 pinned transformers>=4.44,<4.45 → tokenizers==0.19.1).

Security

• Model catalog pinned to immutable Hugging Face commit SHAs under Unplug-AI/; unpublished medium/large tiers removed.

Also since 0.5.0

• Beginner onboarding and agent-host guides; public API facades under unplug.api.*; integrations hub / audit UX polish.

See CHANGELOG.md for full details.

v0.5.0

unplug-ai 0.5.0

Ten agent-framework integrations and a larger security matrix.

Added

• Ten agent-framework integrations: OpenAI Agents SDK, LangChain, Google ADK, smolagents, DSPy, Strands Agents, Letta, Griptape, AG2, and Atomic Agents.
• Per-framework optional extras (unplug-ai[<framework>]) folded into the integrations meta-extra; per-framework guides under integrations/ and runnable demos under examples/.
• Agent security matrix expanded 40 → 72 angles, with dedicated isolated Integrations (live) CI legs per framework.

See CHANGELOG.md for full details.

v0.4.1

Agent-framework integrations.

Added

• Integration hub (unplug.integrations.*): framework-agnostic AgentHooks + adapters for LangGraph, Agno, CrewAI, AutoGen, LlamaIndex, Pydantic AI, Semantic Kernel, and a custom-loop guide
• Per-framework extras (unplug-ai[langgraph], [agno], [crewai], [autogen], [llama-index], [pydantic-ai], [semantic-kernel], [mcp]) + an integrations meta-extra
• 40-angle agent security matrix and live per-framework CI job (tests/optional/live/)
• Integrations docs hub under integrations/

Changed

• all extra installs capability extras only; agent frameworks install via integrations

Install: pip install unplug-ai==0.4.1 · agent stacks: pip install "unplug-ai[integrations]"

unplug-ai 0.4.0

unplug-ai 0.4.0

Minor release bundling audit remediation and new observability APIs.

Added

• ScanResult.degraded and ScanResult.degraded_layers — surface when configured protection layers were unavailable
• Token privacy filter (build_privacy_filter, ModelPrivacyFilter, HeuristicPrivacyFilter) behind optional presidio / ML extras
• MIGRATION.md — API stability tiers, deprecated import paths, and v1.0 removal timeline
• refresh_scan_result stable export at unplug.api.results
• README benchmark table separating regex-only vs ML recall-gate metrics
• with_tiny() recall-gate preset tuned for higher ML recall on injection

Fixed

• Privacy filter thread-safe model load and max_length forwarding
• Benchmark ML guard pipeline isolation and exfil demo output edge cases

Changed

• Flat unplug.core.* shim imports emit DeprecationWarning; canonical subpackages are preferred
• unplug.guard_scan emits deprecation warning — use unplug.api.results
• Judge default model updated to gpt-5.4-nano

Install

pip install "unplug-ai>=0.4.0,<0.5"

Full changelog: CHANGELOG.md

unplug-ai 0.3.1

unplug-ai 0.3.1

Patch release: audit remediation follow-ups, unplug-scan-pr CLI, and GitHub scan distribution.

Install

pip install "unplug-ai>=0.3.1"
pip install "unplug-ai[ml]>=0.3.1"

Highlights

• unplug-scan-pr — scan changed agent/MCP config files in CI
• unplug-scan-action — GitHub Marketplace action (UnplugAI/unplug-scan-action@v1)
• Security fixes: USER secrets default policy, batch_scan error handling, model cache verification

See CHANGELOG.md for full notes.

unplug-ai 0.3.0

unplug-ai 0.3.0

OSS extraction release — decision gating, offline hardening, RAG defense, and eval harness.

Highlights

• DecisionPolicy — unified DOC_OR_SPAN / DOC_ONLY / DOC_GATED modes with regex-to-ML gray-band gating
• Canary tokens — Rebuff-style prompt leak detection in OutputPipeline
• Malicious URL scanner — data URIs, credential-in-URL, IP literals, punycode/homoglyph hosts, shorteners
• Token limits — LimitConfig wired into Guard for OWASP LLM10 input bounds
• Attack harness — garak corpus (51 samples), PyRIT-style converter matrix, make attack-gate CI gate
• Haystack integration — UnplugDocumentGuard component for RAG retrieval defense (pip install unplug-ai[haystack])
• Disposition-aware ML — v132-ready ternary head inference path with harmful-not-injection suppression
• CI — pytest coverage PR comments; integration branch moved to dev

Branching

• PRs target dev (default branch)
• main is updated only on version releases

Install

pip install unplug-ai==0.3.0
pip install "unplug-ai[haystack]==0.3.0"  # RAG integration

unplug-ai 0.2.3

Sensitive-context dual mode, credential-harvest injection patterns, Luhn credit card detection, XSS and path traversal patterns, invisible_text findings, heuristic privacy filter, disposition groundwork. Full diff: v0.2.2...v0.2.3

unplug-ai 0.2.2

unplug-ai 0.2.2

Security fixes

• ABSTAIN policy ordering: high-risk uncertain ML signals BLOCK before abstain
• ScanPolicy.tau_abstain_low wired into injection_ml
• ABSTAIN sentinel span no longer trips coverage gate incorrectly

New modules

• core/disposition.py — injection vs harmful-not-injection placeholder (v132)
• P0 scanner gaps: Luhn credit card, XSS payloads, path traversal, invisible-text finding
• Heuristic privacy filter (dev_heuristic=True)

Install

pip install "unplug-ai==0.2.2"
pip install "unplug-ai[ml]==0.2.2"