AgentArmor v1.4.3

AgentArmor v1.4.0 — Multi-agentic only

AgentArmor is now a multi-agent AI red-teaming platform. The offline analysis mode has
been removed: every scan runs capability-aware attack-graph planning, an LLM judge, and
confidence scoring. This release also fixes real correctness bugs in the multi-agentic engine.

Changed

• Offline analysis mode removed — the GUI offline↔cloud picker is gone; the CLI and API
  run multi-agent (cloud) analysis on every scan.
• Analysis API key now required — scans without a provider key are rejected with a clear
  error. Set AGENTARMOR_ANALYSIS_API_KEY, or configure it in Settings / AgentArmor.toml.
• The local L1–L4/meta detection engine still runs as the under-the-hood scorer; local
  model-file scanning (--model) is unaffected.

Fixed (multi-agentic engine)

• Module targets now receive the real attack — red-team attacks against agent/MCP/RAG
  targets send the generated attack (harness prompt, RAG query, MCP tool-parameter injection)
  instead of running a static probe and relabelling the display text.
• Budget metering — the LLM judge and web attack planner now record token/cost usage
  against the budget governor instead of bypassing it.

Added (multi-agentic engine)

• Per-path campaign coverage — a finding now retires only its attack path and the campaign
  continues across the remaining paths, instead of stopping at the first finding.

---

Windows (recommended)

Download AgentArmor_1.4.0_x64-setup.exe or AgentArmor_1.4.0_x64_en-US.msi from the assets below.

1. Run the installer
2. Open AgentArmor — no Python install required
3. Add your analysis provider API key in Settings
4. Choose a scan profile → run scan → review grouped findings → download reports

AgentArmor v1.4.2

AgentArmor v1.4.0 — Multi-agentic only

AgentArmor is now a multi-agent AI red-teaming platform. The offline analysis mode has
been removed: every scan runs capability-aware attack-graph planning, an LLM judge, and
confidence scoring. This release also fixes real correctness bugs in the multi-agentic engine.

Changed

• Offline analysis mode removed — the GUI offline↔cloud picker is gone; the CLI and API
  run multi-agent (cloud) analysis on every scan.
• Analysis API key now required — scans without a provider key are rejected with a clear
  error. Set AGENTARMOR_ANALYSIS_API_KEY, or configure it in Settings / AgentArmor.toml.
• The local L1–L4/meta detection engine still runs as the under-the-hood scorer; local
  model-file scanning (--model) is unaffected.

Fixed (multi-agentic engine)

• Module targets now receive the real attack — red-team attacks against agent/MCP/RAG
  targets send the generated attack (harness prompt, RAG query, MCP tool-parameter injection)
  instead of running a static probe and relabelling the display text.
• Budget metering — the LLM judge and web attack planner now record token/cost usage
  against the budget governor instead of bypassing it.

Added (multi-agentic engine)

• Per-path campaign coverage — a finding now retires only its attack path and the campaign
  continues across the remaining paths, instead of stopping at the first finding.

---

Windows (recommended)

Download AgentArmor_1.4.0_x64-setup.exe or AgentArmor_1.4.0_x64_en-US.msi from the assets below.

1. Run the installer
2. Open AgentArmor — no Python install required
3. Add your analysis provider API key in Settings
4. Choose a scan profile → run scan → review grouped findings → download reports

AgentArmor v1.4.1

AgentArmor v1.4.1 — Multi-agentic only

AgentArmor is now a multi-agent AI red-teaming platform. The offline analysis mode has
been removed: every scan runs capability-aware attack-graph planning, an LLM judge, and
confidence scoring. This release also fixes real correctness bugs in the multi-agentic engine.

Changed

• Offline analysis mode removed — the GUI offline↔cloud picker is gone; the CLI and API
  run multi-agent (cloud) analysis on every scan.
• Analysis API key now required — scans without a provider key are rejected with a clear
  error. Set AGENTARMOR_ANALYSIS_API_KEY, or configure it in Settings / AgentArmor.toml.
• The local L1–L4/meta detection engine still runs as the under-the-hood scorer; local
  model-file scanning (--model) is unaffected.

Fixed (multi-agentic engine)

• Module targets now receive the real attack — red-team attacks against agent/MCP/RAG
  targets send the generated attack (harness prompt, RAG query, MCP tool-parameter injection)
  instead of running a static probe and relabelling the display text.
• Budget metering — the LLM judge and web attack planner now record token/cost usage
  against the budget governor instead of bypassing it.

Added (multi-agentic engine)

• Per-path campaign coverage — a finding now retires only its attack path and the campaign
  continues across the remaining paths, instead of stopping at the first finding.

---

Windows (recommended)

Download AgentArmor_1.4.1_x64-setup.exe or AgentArmor_1.4.0_x64_en-US.msi from the assets below.

1. Run the installer
2. Open AgentArmor — no Python install required
3. Add your analysis provider API key in Settings
4. Choose a scan profile → run scan → review grouped findings → download reports

AgentArmor v1.4.0

AgentArmor v1.4.0 — Multi-agentic only

AgentArmor is now a multi-agent AI red-teaming platform. The offline analysis mode has
been removed: every scan runs capability-aware attack-graph planning, an LLM judge, and
confidence scoring. This release also fixes real correctness bugs in the multi-agentic engine.

Changed

• Offline analysis mode removed — the GUI offline↔cloud picker is gone; the CLI and API
  run multi-agent (cloud) analysis on every scan.
• Analysis API key now required — scans without a provider key are rejected with a clear
  error. Set AGENTARMOR_ANALYSIS_API_KEY, or configure it in Settings / AgentArmor.toml.
• The local L1–L4/meta detection engine still runs as the under-the-hood scorer; local
  model-file scanning (--model) is unaffected.

Fixed (multi-agentic engine)

• Module targets now receive the real attack — red-team attacks against agent/MCP/RAG
  targets send the generated attack (harness prompt, RAG query, MCP tool-parameter injection)
  instead of running a static probe and relabelling the display text.
• Budget metering — the LLM judge and web attack planner now record token/cost usage
  against the budget governor instead of bypassing it.

Added (multi-agentic engine)

• Per-path campaign coverage — a finding now retires only its attack path and the campaign
  continues across the remaining paths, instead of stopping at the first finding.

---

Windows (recommended)

Download AgentArmor_1.4.0_x64-setup.exe or AgentArmor_1.4.0_x64_en-US.msi from the assets below.

1. Run the installer
2. Open AgentArmor — no Python install required
3. Add your analysis provider API key in Settings
4. Choose a scan profile → run scan → review grouped findings → download reports

AgentArmor v1.3.1

AgentArmor v1.3.1 — Detection stack overhaul

Release focused on detection accuracy, reproducibility, and enterprise-ready policy hooks.

Added

• Echo-aware scoring — L1/L2 echo stripping; tiered compliance assertions (Sprint 1)
• ONNX/FAISS hardening — tokenizer bundles, index versioning, honest fallback (Sprint 2)
• Regression harness — 70 fixtures, agentarmor eval detection, baseline JSON (Sprint 2)
• Unified judge — JudgeService with legacy config migration (Sprint 3)
• Rule catalog — single YAML source for L1/L2/L4 (Sprint 3)
• Per-probe thresholds + meta calibration scaffold (Sprint 4)
• Detector plugins — SDK, marketplace --trust, LLM-rubric assertions (Sprint 5)
• Webscan partial-stream gate — completeness-aware WARN/FAIL (Sprint 5)
• Policy engine, evidence spans, version stamps, active-learning queue (P5)

Fixed

• Meta scorer hard-signal floor for clear L1/L4 hits
• Webscan partial streams no longer blanket-WARN when hard outcomes are present

---

Windows (recommended)

Download AgentArmor_1.3.1_x64-setup.exe or AgentArmor_1.3.1_x64_en-US.msi from the assets below.

1. Run the installer
2. Open AgentArmor — no Python install required
3. Choose a scan profile → run scan → review grouped findings → download reports

AgentArmor v1.3.0

AgentArmor v1.3.0 — Scan quality overhaul

Major release: OWASP Planner v2, finding clustering, calibrated progress/ETA, scan profiles, and production hardening.

Added

• OWASP Planner v2 — ~50 probes at standard depth, Quick/Standard/Deep budgets, capability-aware selection
• Finding clustering — root-cause cards, semantic merge, replay bundles, grouped findings API
• Scan progress — weighted work units, ETA confidence bands, planner-aware layer tracking
• Risk-based planning — probe reorder after early failures; adaptive depth escalation
• Scan profiles — 8 presets including production_readiness and owasp_audit (GUI picker)
• API — plan preview, metrics, regression compare, attack narrative graph
• Parallel L1/L2 batches with per-probe fault isolation

Fixed

• Failed scans set FAILED status (no stuck RUNNING)
• Budget preflight rejects over-limit scans before start
• Grouped findings endpoint and web planner crashes

---

Windows (recommended)

Download AgentArmor_1.3.0_x64-setup.exe or AgentArmor_1.3.0_x64_en-US.msi from the assets below.

1. Run the installer
2. Open AgentArmor — no Python install required
3. Choose a scan profile → run scan → review grouped findings → download reports

AgentArmor v1.2.4

AgentArmor v1.2.4 — Readable findings, scan progress, and report download

Latest release. Makes scan results easier to understand, adds live progress with ETA, and one-click report downloads from the GUI.

Added — GUI UX

• Findings: executive summary, excerpt-first cards sorted by severity, attack chains collapsed by default
• Scan progress: live stopwatch, progress bar, ETA, human probe labels; technical SSE log collapsed
• Reports: download PDF, HTML, SARIF, JSON, or ZIP from Findings, Progress, and Reports pages
• Report download API with path traversal guard (GET /v1/scans/{id}/reports/download)

Fixed

• composite_vuln_score() crash when all assertions pass (agent scans)
• Empty Authorization: Bearer header from config breaking connectivity checks

---

Windows (recommended)

Download AgentArmor_1.2.4_x64-setup.exe or AgentArmor_1.2.4_x64_en-US.msi from the assets below.

1. Run the installer
2. Open AgentArmor — no Python install required
3. Run a scan → review findings with plain-language summaries → download reports from the header

For login-required chatbot targets, use Login required (SSO) in the website wizard.

AgentArmor v1.2.3

AgentArmor v1.2.3 — Packaging and docs

Patch release aligning Windows installer filenames with the app version and updating release documentation.

Fixed

• Tauri bundle version synced to 1.2.3 (AgentArmor_1.2.3_x64-setup.exe / .msi)
• GUI build typing fix for web-scan status polling

Includes v1.2.2

• Website URL scan SSE reliability and Playwright response capture improvements
• GUI disclaimer on website URL vs API endpoint testing limits

---

Windows (recommended)

Download AgentArmor_1.2.3_x64-setup.exe or AgentArmor_1.2.3_x64_en-US.msi from the assets below.

1. Run the installer
2. Open AgentArmor — no Python install required
3. Test my chatbot → website URL or API URL → run scan → export reports

For login-required targets (ChatGPT, Claude), use Login required (SSO) in the website wizard.

AgentArmor v1.2.2

AgentArmor v1.2.2 — Website URL chatbot scanning

Latest release. Fixes long-running website URL scans disconnecting from live progress, and improves Playwright discovery and response capture for real chatbot UIs.

Fixed — Website URL chatbot scanning

• SSE scan progress: 30s heartbeats and discovery/planning events so long web scans no longer show "Connection to scan stream lost"
• Scan progress UI falls back to status polling if the live stream disconnects
• Playwright response capture uses baseline diffing and chat-root detection (Gandalf, Prompt Airlines-style UIs)
• Send-button heuristics skip upload/attach controls; async route guard fix in browser session
• SPA page wait for visible chat inputs; webscan.timeout_s default raised to 60s
• Low-confidence widget fallback and optional llm_discovery_on_miss config
• ScanSummary metadata typing for web-scan status polling (GUI production build)

Added

• Framework hints for ChatGPT, Claude, and Gemini; chat keywords for challenge sites
• Gandalf-like HTML fixture and Playwright regression test
• GUI disclaimer on website URL testing limits vs API endpoint scanning

---

Windows (recommended)

Download AgentArmor_1.2.1_x64-setup.exe or AgentArmor_1.2.1_x64_en-US.msi from the assets below (v1.2.2 app; installer filenames use 1.2.1).

1. Run the installer
2. Open AgentArmor — no Python install required
3. Test my chatbot → website URL or API URL → run scan → export reports

For login-required targets (ChatGPT, Claude), use Login required (SSO) in the website wizard.

AgentArmor v1.2.1

AgentArmor v1.2.1 — Multi-agent red team & web scan

This release ships the multi-agent OWASP red team roster and enterprise chatbot URL scanner improvements from the latest main branch.

Multi-agent OWASP red team

• agentarmor/redteam/ orchestrator with attack-graph planner, budget governor, and verdict scoring
• 16 skill YAML bundles and 13 dedicated agent classes (LLM01–LLM10, Memory, A2A, MCP)
• scan_mode=multi_agent_redteam on /v1/scans (requires cloud analysis API key)
• Web scans: multi_agentic + planner_enabled runs red team against discovered widgets
• Findings include confidence, reproducibility, and impact scores
• GUI toggle for multi-agent red team; quality gates and regression tests in CI

Chatbot URL scanner

• Enterprise manual_session auth with headed browser login and encrypted session storage
• POST /v1/web-scans/prepare-session and continue flow for SSO-style targets
• LLM attack planner for multi-agentic web scans
• Multi-turn memory probe without page reload
• Daily web scan rate limiting and HTML report attack-plan summary

Packaging

• Embedded Python bundle includes red team skills and Playwright for packaged .exe builds
• Version aligned to 1.2.1 across Python, Tauri, and npm

---

Windows (recommended)

Download AgentArmor_*_x64-setup.exe or .msi from the assets below.

1. Run the installer (or portable .exe)
2. Double-click AgentArmor — no Python install required
3. Pick a scan type → configure → run → export HTML/SARIF/PDF from Reports

Portable mode: run from a folder containing a PORTABLE file; data is stored in ./data/.

macOS

A native Mac .app is not bundled (Windows desktop first). Use the CLI + web UI:

pip install agentarmor
pip install agentarmor[local]   # optional: offline .gguf scanning
agentarmor models download
agentarmor serve --port 8787
cd AgentArmor/gui && npm install && npm run dev

See docs/MAC.md.

Linux / PyPI

pip install agentarmor
agentarmor scan --url http://localhost:8000/v1/chat/completions

GitHub Action (CI)

- uses: YashvantHange/AgentArmor/action@v1
  with:
    url: https://api.example.com/v1/chat/completions