AI Security Matrix

A curated list of AI-enabled security testing tools.

Week of 9 to 16 November 2026

29 of the 83 tools shipped changes last week, across 1549 commits and 26 releases. None left.

Last Week's Highlights

  • google/mantis: A new RemoteTargetEnvironment runs split-plane dynamic testing against a full running product.
  • microsoft/PyRIT: PyRIT can now pull trace evidence from a remote OTLP/JSON backend and score whether tool calls occurred in the expected order.
  • MythicAgents/Sage: The Supervisor can propose an explicit BloodHound collection, including one aimed at a missing objective domain, and runs it as Sage's own procedure once requested.

AIPentest/CyberStrikeAI

v1.8.0, 38 commits

CyberStrikeAI's agent runtime is now Pi, replacing Eino as the sole execution engine, with native session storage and steering. Pi reduces DeepSeek input overhead and auto-detects local model context size. The vulnerabilities view now groups findings into a review workspace with retest tracking, timestamps, and an oldest-first sort order. Pi skill previews now count tool calls correctly. Deleted sessions are now invalidated instead of persisting. One-click updates switch versions without interrupting an active sign-in, activate new binaries without a supervisor process, and reuse the same HTTPS certificate across updates.

Armur-Ai/Pentest-Swarm-AI

21 commits

Pentest-Swarm-AI now computes a fix-plan and cut-set path analysis across the attack graph, showing the minimal fix set on the HTML attack map and per-step fallback routes to the objective. The report surfaces that fix plan, carries its leverage into SARIF results, and adds a Chain Coverage section showing which attack chains were considered versus proven. The executive summary gets a deterministic "At a glance" headline, and the Remediation Plan now orders findings by blast radius before severity. AssessmentSnapshot exports now carry unique edge ids and are validated against a DiscoveryBundle before a warm-start. Scope checks now validate the host a target points at, not its path. The --budget spend cap is enforced on the sequential (default) path.

KeygraphHQ/shannon

v3.5.1, 3 commits

Shannon adds a documentation-backed skill plugin, giving testers a packaged way to run Shannon's capabilities through that plugin interface. Git and subprocess output now gets decoded once instead of multiple times, so multi-byte characters in findings and logs no longer get garbled. Workspace names are now validated and target URLs are restricted to HTTP/HTTPS. The OpenAI cyber-access error message now links to chatgpt.com/cyber.

Lab700xOrg/aisbom

4 commits

The comply command now fills TODO fields from a provider answers file, and pre-fills more fields directly from scan data. Scanning now times out stalled Hugging Face requests instead of hanging, and page URLs that previously failed silently are now reported as errors. The GitHub Action adds a fail-on-scan-error input so a scan failure stops the run instead of letting it report a stale SBOM.

MythicAgents/Sage

v0.3.0, 13 commits

Sage now checks supervised delegations against their assignment and sources before letting a step proceed. The Supervisor can propose an explicit BloodHound collection, including one aimed at a missing objective domain, and runs it as Sage's own procedure once requested. The supervised path now closes end to end, from graph collection through forge and cleanup. Sage learns each domain's FQDN from the operation's own evidence rather than needing it supplied. A follow-up message can now name the target the operator gave earlier in the same chat. Sage refreshes the callback's Kerberos context before an admin-backed DCSync so that step does not run on stale credentials.

NVIDIA/SkillSpector

v2.13.5, 229 commits

SkillSpector now scans skills through a MiniMax regional provider with M3.1 Flash as the default model. The nv_build default model switches to z-ai/glm-5.3-flash. Behavioral analysis gains a canonical dangerous-callable resolver backed by a shared evasion corpus. The CLI can apply shipped baselines directly from local ZIP inputs. MCP scan responses can skip rendered reports when you don't need them. The MCP registry now fetches named servers directly. Semantic intent findings are now grounded in observed operations rather than surfaced on their own. Child-process environment pass-through now reports as LOW instead of HIGH, and scanner-budget gaps no longer score as HIGH AE1 findings.

PurpleAILAB/Decepticon

100 commits

Decepticon can now schedule engagement runs with a LangGraph cron schedule and pause and auto-resume runs across a defined time window. The launcher brings up an OpenShell gateway on start, choosing docker or podman automatically. The scope wizard reuses a postgres-backed asset inventory and enforces optional RoE controls before an engagement moves from draft to started. Sandbox providers are now replaceable, with the former http provider renamed docker, and browser automation runs through replaceable providers with Obscura as the default. Scope is now enforced only through OpenShell. The open-web engine no longer accepts a non-2xx JSON response as a successful fetch, and now works when run under OpenShell. The OSS web dashboard has been removed; use the SaaS UI instead.

SnailSploit/Claude-Red

4 commits

Claude-Red corrects several techniques in the Skills library that had been verified but contained errors, so testers relying on those entries now get accurate guidance. The Skills index has been regenerated to match the current set of skills after those corrections. Nothing else changed in the tool last week.

Tencent/AI-Infra-Guard

v4.6.10, 22 commits

AI-Infra-Guard adds a local generative AI safety self-check as a new compliance check. The fingerprint scanner now recognizes one-api deployments. The CVE dataset gains advisories for AI products pulled from NVD. mcp-scan, skill-scan and agent-scan now keep findings that previously got dropped: hex-suffixed tool-call tags are normalized, and missing level or suggestion tags no longer drop a finding. Batch arrays decode from XML parameters, prompt escapes survive provider templates, and non-object HTTP error responses are preserved instead of discarded. skill-scan retries transient LLM errors instead of failing the whole scan. Scanners now forward reasoning effort to the model API, and httpx honors configured Host headers.

berylliumsec/nebula

nebula-v3.0.46, more than 300 commits

Nebula's terminal client now matches the GUI, with a Work board at /board, a project menu, goal, approval and subagent controls, model choice, chat lifecycle actions, and a follow up queue, so you can run and manage a full project from the terminal alone. The terminal's project picker now searches the complete project catalog with multi-word, typed-ahead search. Nebula removed its built-in attack workflows and scanner approval demos, leaving only browser navigation primitives to build tests on. Execution workers now run multiple projects side by side instead of one at a time. The iPhone app's build 6 bundles Butler and a Projects snapshot. A host pressure threshold of 100% now turns off the pressure check entirely.

blacklanternsecurity/red-run

1 commit

The repository has been archived. No functional changes shipped last week, only a README update. If you run red-run today, expect no further updates or fixes from this project going forward.

elementalsouls/Claude-BugHunter

53 commits

A new coverage ledger tracks engagement observations and a closure gate now blocks a run from finishing until coverage is complete. Scope matching preserves metacharacters in host regex rules, rejects malformed target URLs without aborting the run, and rechecks the current scope before resuming validation. The engine resumes validation by parameter identity, isolates individual worker failures during serial runs, and persists positive hunt results atomically. The CLI retains recon headers regardless of casing and reports incompatible recon manifest shapes. Installation discovers Python executables in PowerShell and uninstalls only from tracked harnesses, preserving unrelated profile lines.

google/mantis

26 commits

Mantis now accepts --depth 1..100 to control scan depth, with new repo docs covering a cross-domain checklist. The surveyor and index add cross-domain and RTL support plus 13 additional grammars, widening what source trees it can parse. A new RemoteTargetEnvironment runs split-plane dynamic testing against a full running product. Mantis now runs a post-run cross-campaign semantic deduplication pass and produces an aggregate report across campaigns. The audit planner and researcher seed from a deterministic AST sink and entrypoint checklist, configurable with a --passes count. Directory-slice campaigns are now anchored to their assigned target and verify reads stay in-slice. Resume now recovers from NodeContextExhausted and provider context overflows through a structured handoff. Context editing now reads the MANTIS_MAX_INPUT_TOKENS environment variable.

h5i-dev/h5i

v0.5.5 to v0.5.6, 16 commits

The HTTP head parser now enforces stricter validation of the request target and the Host header, proven sound under an independent specification. Proofs now also cover response heads, whole-body decoding, and message boundaries, so requests and responses captured through the browser proxy are parsed to their correct length and split points rather than approximated.

microsoft/PyRIT

v1.2.0, 153 commits

PyRIT can now pull trace evidence from a remote OTLP/JSON backend and score whether tool calls occurred in the expected order. ToolsCalled can assert that a specific set of tools ran and nothing else, and check tool arguments such as send_email to=x. New scenarios cover aggregation leakage across permitted documents, Garak's LeakReplay training-data replay, and an ASSERT policy scenario with judge dimensions and presets. SecretDisclosureScorer flags partial and obfuscated disclosure of a known secret. CredentialLeakScorer now detects OpenPGP private key blocks, encrypted private keys, and GitLab tokens. A prefill technique adds native assistant continuation. TAP runs now report ERROR when every branch crashes before scoring.

mpgn/NetExec-mcp

v1.0.9 to v1.0.10, 15 commits

Search now recognizes s4u, process, and golden-ticket intents, with guidance on forging tickets. Two new tools, smb_klist and winrm_klist, list or dump the TGTs cached on a host, matching dumped tickets against a before/after snapshot and reporting klist's own failures. Get-file tools now anchor their destination path and report what ends up on disk. The verdict now marks an admin SSH login as pwned and reads ccache, certificate, and S4U logins as logins. Dynamic argument handling coerces near-miss list arguments and names the real expected type in validation hints. Input files are read from wherever the tools actually write them, and NXC_PATH is pinned at startup.

mrexodia/ida-pro-mcp

2 commits

fix: declare_type now uses default struct packing. Nothing else changed in the tool.

msoedov/agentic_security

23 commits

Agentic Security tags probes with their OWASP LLM Top 10 2025 ids. The harness no longer misclassifies blank-line outcomes, so run results reflect actual pass or fail states. JSON control characters in the PROMPT substitution are now escaped, fixing broken payloads in http spec tests. The tool adds Requesty and UberLLM presets to the LLM specs, letting you target either provider directly from the spec list.

promptfoo/modelaudit

16 commits

modelaudit now flags UUIDv4 credentials found in explicit Bearer headers as a secret. Text scanning reports README examples and public IP addresses as informational findings instead of risk findings, cutting false positives. Skops scanning does the same for README loader references. Pytorch scanning attributes verified legacy TorchScript controls to the native loader rather than misreporting them. Manifest scanning identifies Hugging Face download records specifically. Picklescan qualifies frame warnings that come from encoded probes, and license checks now qualify rights notices for pickle files. ONNX weight hashing no longer misroutes large files through Flax.

promptfoo/promptfoo

64 commits

Promptfoo adds support for the Gemini Decisions API, giving testers a new provider target. Gemini 4 Argon and Hydrogen and Claude Fable and Mythos 5.2 are now supported providers for eval runs. Tool call history is preserved across multi-turn conversations with Gemini and Anthropic providers, so evals no longer lose context mid-run. Vertex Claude and Llama providers now preserve empty completions. Measured provider latency is now reported in whole milliseconds and never reported as negative. Vertex Llama models now receive generationConfig sampling options. GPT-6.1 Terra, GPT-6 Sol, and GPT-6 Astra are now priced correctly on Bedrock Converse. The init flow no longer offers the retired vertex:gemini-2.5-pro model.

samugit83/redamon

v7.8.0, more than 300 commits

RedAmon runs as a multi-user tool now, with a confirmed Root, 98 named permissions, and project sharing at Read or Read plus Write. You share a project from its card or header to an exact email as Viewer or Analyst, with an optional end date, and the recipient sees it under Shared with me. A collaborator at Read reaches the owner's graph, chains, reports and schedules without ever seeing the owner's tokens, RoE document or custom headers. Read plus Write adds conversation sharing, consent to spend the owner's credentials, and ownership transfer confirmed by both passwords. Only one person writes a shared project at a time, and a second scan or agent run is refused and names who holds it. The IAM console, opened from the header where Users used to be, lists users, permission sets, shares, quotas and instance LLM providers at /admin. AUDIT_FORWARD=stdout prints every audit row as a JSON line for your log pipeline. The five RBAC switches ship on by default, so if you script setup, confirm your .env sets them explicitly or the update turns them on for you.

snyk/agent-scan

7 commits

Agent-scan now retries hooks automatically when a test event fails with a transient error, so a flaky connection no longer fails the whole run. A certificate error that could interrupt scans has been fixed. Free-tier users are now capped at a daily usage limit.

splx-ai/agentic-radar

v0.14.2, 2 commits

Agentic Radar now loads environment variables only from a path you give with --env-file, instead of picking up a .env file on its own. If you script it, pass --env-file explicitly to load environment variables.

trailofbits/fickling

1 commit

Two dependency bumps. Nothing changed in the tool.

trailofbits/skills

1 commit

Trailmark adds a taint-trace skill to the collection, letting you trace tainted data through a codebase as part of a security review. The skill runs through a native analyzer workflow rather than relying on general-purpose code reading, giving taint-trace results you can call directly in an agent session.

usestrix/strix

v1.7.1, 15 commits

Strix sharpens its security impact-chain detection skill and adds a check for query-language and JSON parser differentials. Dependency reports now carry one reachability verdict per CVE, drop a leftover static-analysis disclaimer from the evidence, and accept code_locations so findings point at the relevant file. MCP connections now refresh their bearer token on a 401. The sandbox image pins katana to v1.8.0 and installs gitleaks once, and Go tools are rebuilt with patched dependencies. Windows connections launched through go run get a longer handshake window, and the startup limit now applies to them correctly.

vestjoe/trove-skill

26 commits

Trove-skill now reads Hydra output as confirmed logins on their host and port, kerbrute output as principals and confirmed Kerberos logins, and John --show output as cracks joined by hash or by account and realm. AI extraction now names each finding by the quoted passage from its title and host, sends each chunk to the model only once, and keeps notes only when they come from the user's own words. The tool documents that kerbrute v1.0.3's -o file drops the expired-password reason, and now expects that account directly from a kerbrute userenum capture. NetExec --log import holds back a command line when the log already carries a login line. Import --realm now tags a user list with its realm.

xalgorix/xalgorix

v4.6.185 to v4.6.196, 33 commits

Single-host scans now keep requests off other organizations' domains found in pages, redirects or JavaScript, scanning only the target's own subdomains. The scope guard no longer blocks interactsh callback servers, SSRF placeholders, metadata hostnames, egress checks and vulnerability databases, cutting dozens of false blocks. Every HTTP finding now ships a replayable curl command built from its method and endpoint, with the method shell-quoted so the line runs as one command. http_request and send_request now check each redirect hop before following it, refusing one that points at the operator's machine or the dashboard listener. Wildcard scans now resume a host interrupted by a server restart from its saved session instead of skipping it, and no longer log a phantom sub-scan.

zhaoxuya520/reverse-skill

61 commits

Reverse-skill now triages stripped Go and Rust binaries as a target type. The IDA debugger integration can now auto-patch binaries during a session. Routing fixes resolve a tie so R14 LLM red-team keyword rules fire instead of being blocked by R10, and anchor the R5 .net keyword so FQDNs no longer trigger false positives. CJK text handling now resolves word boundaries between adjacent characters correctly. Radare2 recon no longer aborts on a SIGPIPE when its output is truncated. The tool ships as an all-in-one Docker image, and a new Orca setup hook simplifies installation.

What is a highlight?

Each week an AI model reads every entry and picks up to three changes that matter most to the people running these tools. It looks for these, most important first. Each highlight quotes the sentence from the tool's entry that shows the change, with the sentence before it when it needs one.

1 The tool can run a test, technique or attack it could not before, or reach a new kind of target.
2 The tool produces a new kind of output a tester can hand on, such as an attack map, a report or verified proof of a finding.
3 The tool's results can be trusted where they could not before: findings are now verified, or a fix corrected results that were wrong.
4 The tool tests more of a target than before: more vulnerability classes, more hosts or roots, more of a methodology.
5 The tool keeps a test inside its scope or rules of engagement in a way it could not before.