beenuar/AiSOC

This page introduces an external open-source repository. It is not an HDATF product.

What it does

AiSOC is a self-hosted security operations platform that consolidates alerts and records AI investigation evidence and proposed responses. It takes events from security tools connected by operators and produces alerts and investigation records through normalization, detection and correlation. Kafka transports events and PostgreSQL stores alerts and audit records, with results exposed in a web console. Prompts, tool calls and verdict evidence are recorded in the Investigation Ledger. The default deployment uses a local model without an external model API key, but real company telemetry requires separate connector setup. Its documented base environment needs Docker Compose v2, 8 GB memory and 20 GB free disk in the Docker VM, plus Python and bash. Sample data and the offline sandbox must be distinguished from real incidents and the complete production stack. Response is human-approval-oriented by default, with explicit policy authorization required for autonomy. Official material still discloses limits in hosted-model testing and sustained-operation evidence, so it does not establish every integration as verified. It is not a drop-in replacement for SIEM long-term retention and compliance search.

License

MIT Permissive. Commercial use and changes are allowed if the copyright notice is kept.

More in this category

• onlyreadai/onlyread
  OnlyRead is a Chromium extension intended to reduce accidental writes when an AI agent browses real signed-in work accounts. Its input is outgoing traffic to guarded sites, and its output is a decision to continue or refuse each request rather than a rewritten page or a content summary. The inspected decision code combines method and endpoint rules, site-specific operation decoding, reviewed request labels and model judgment for remaining cases. Chrome's debugger interface pauses requests, with additional browser network rules and site-specific controls for some socket and editor paths. Setup requires a compatible Chromium browser, selected guarded sites and either the hosted judge account or credentials for a supported model gateway. Uncertain or unavailable judgment is designed to refuse a request, but the model can still misclassify traffic and the system allows policy-specific exceptions such as telemetry. Screen-driving agents need the Always setting, protection is currently site-wide rather than account-specific, and the architecture documents paths where editor guards are not network guarantees. Some redacted request descriptions leave the browser for judgment, and this review did not independently reproduce the repository's site coverage or security claims.
• Verlintas/NOIR
  NOIR is a desktop tool for inspecting privacy risks on Android devices that the operator owns or is authorized to audit. It reads installed applications, permissions and network address strings in APK files through adb, the Android device connection tool. It associates permission categories with possible destination domains and produces risk scores, JSON and Markdown reports, domain lists and remediation plans. Shared user permissions and application manifests provide additional evidence, while optional deeper inspection uses the separate local jadx decompiler. Its recommendations distinguish reversible application permission controls from actions requiring manual review and can produce rollback commands. The documented starting point is a portable package for Windows, macOS or Linux and a device that has authorized USB debugging, with adb prepared separately. Portable packages include a Java runtime, while source execution requires JDK 17 or later, and tool setup can download missing dependencies. These are static associations rather than proof of an actual transmission, and restricting permissions still requires checking whether legitimate application functions are affected.
• interplains/Zero-Breach-Toolkit
  Zero Breach Toolkit is a defensive local checker that compares device evidence with published spyware indicators. It targets people examining their computers or phone backups for known traces. Inputs include browser records, selected operating-system artifacts, file fingerprints, AndroidQF collections and local iPhone evidence. Readable and JSON reports show matches, evidence strength and coverage gaps. Ordinary scans preserve source evidence, use temporary database copies and block Python socket calls. That guard is not operating-system network isolation, and indicator updates are a separate online action. Basic execution uses Python 3.8 or newer, while encrypted iPhone backups require an additional cryptography package and its supported environment. Optional Android acquisition requires owner consent and can change device state, unlike read-only scanning. No matches do not prove safety, and matches do not automatically establish infection. Full reports and unencrypted evidence ZIPs can contain sensitive records and need restricted sharing.
• apple/security-pcc
  Apple publishes selected Private Cloud Compute components so researchers can examine its security and privacy properties. This is not a ready-to-install chatbot or general inference-server product. The repository contains security-policy components and research tools and points readers to an accompanying official security guide. The inspected attestation-verification CLI accepts a report or attestation-bundle file and dispatches it to a format-specific validator. Its output paths include verification summaries and details or validated key and expiry information. Certificate and transparency-proof settings are present, but inspecting them neither reproduces all PCC nor independently proves its security. Research-tool environments and software terms must be checked before use; cloning does not activate a cloud service. A custom license restricts use to internal PCC verification on one Apple computer for a limited period and prohibits redistribution, unlike ordinary reusable open-source licensing. The guide body was inaccessible in this retrieval, so this description is limited to inspected repository documents and source.
• anthropics/sandbox-runtime
  Anthropic Sandbox Runtime is a CLI and library that restricts file access and networking for arbitrary processes, including agents and local MCP servers. It accepts a command plus filesystem and network policies and executes the command within operating-system boundaries. macOS uses sandbox-exec, Linux uses bubblewrap, and HTTP and SOCKS5 proxies check permitted network destinations. The current README also describes alpha Windows support using a separate account, restricted tokens, file ACLs and WFP network filters. The package requires Node.js 22.12.0 or later, and restrictions apply only when commands are explicitly wrapped or the library is integrated. Linux requires bubblewrap, socat and ripgrep, while Windows requires a separate elevated setup step. Default reads are allowed, so automatic protection of secret files must not be assumed. Read-denied paths, write-allowed paths and allowed domains must be configured for the workload. This is an early research preview, and documented Windows gaps involving alternative tokens and service-mediated traffic prevent treating it as an absolute security boundary.

Only repositories in the ranked Trendshift lists are included, and the lists are used only to find candidates. We do not copy their ranks. Descriptions, licenses and star counts come from each GitHub repository. The notes are our own reading. We have not tested these projects, and a place on a trending list does not prove quality.