Loading story…
Kali Linux published a new installment in its LLM-driven workflow series showing how to run LLM-assisted penetration testing fully offline/on-prem by hosting the model, context/tooling layer, and client locally. The walkthrough uses Ollama (a local model runner) plus 5ire as a desktop client, emphasizing operational security and privacy benefits versus cloud-hosted AI services, and noting the primary tradeoff is GPU hardware and power cost rather than SaaS subscriptions.
The guide demonstrates an NVIDIA CUDA-based setup on a GeForce GTX 1060 (6GB VRAM), including replacing the open-source nouveau driver with proprietary non-free NVIDIA drivers to enable CUDA acceleration and validating the environment with nvidia-smi after reboot. In parallel, SentinelOne published separate research on using LLMs for cyber threat intelligence (CTI) information extraction, describing how models can convert narrative reports into structured data (e.g., IOCs and relationships) and discussing evaluation tradeoffs; this is adjacent AI-in-security content but not part of the Kali local pentesting stack or its implementation details.
Get the actors, campaigns, and ATT&CK mapping behind it.
2 events from the most recent confirmed update back to the earliest known activity.
As part of the published workflow, Kali demonstrated end-to-end natural-language control of Kali tools by having the local LLM stack invoke an nmap scan against scanme.nmap.org through the MCP-connected tool server. The example validated that model inference and tool execution could be performed fully locally.
The Kali Linux team published a guide showing how to run an AI-assisted penetration testing workflow entirely on local hardware using Ollama, mcp-kali-server, and the 5ire client. The guide emphasizes offline operation, NVIDIA GPU acceleration, and avoiding third-party cloud services for privacy and operational security.
Vulnerabilities, threat actors, malware, products, organizations, and breaches Mallory has linked to this story.
Get the adversaries, campaigns, and ATT&CK mapping behind this technique, with detections ready to deploy.
2 references tracked. Mallory keeps watching after this page renders.
Map indicators from this story to your assets and identify affected systems in minutes.
Every observed campaign, victim, and pivot linked to actors named in this story.
Malware, exploits, and IOCs connected to the activity described here.
YARA, Sigma, and Snort rules deployed to your SIEM as soon as they’re published.
Get matching new stories delivered to your team as they break — not the next morning.
Ask questions about this story and take action on the answers.
Other stories Mallory is tracking that overlap on entities, sources, or topic labels.
Kali Linux published documentation for an **AI-assisted penetration testing workflow** that uses Anthropic’s **Claude Desktop** as a natural-language interface and the open-source **Model Context Protocol (MCP)** to translate prompts into tool execution on a Kali host. The described architecture separates a GUI “UI layer” (Claude Desktop on macOS/Windows), an “execution layer” (a Kali system running an `mcp-kali-server` bridge that exposes tools), and an “intelligence layer” (Anthropic’s *Claude Sonnet 4.5* in the cloud). The goal is to let operators request tasks like port scanning and web checks in plain language, with the LLM planning and iterating through tool calls while returning structured results. This is **not an incident report** but a capability/workflow enablement that can materially change how offensive tooling is orchestrated, including by lowering the operational friction of chaining common recon and enumeration steps. Organizations evaluating similar agentic workflows should treat the MCP bridge and any AI agent runtime as a privileged automation surface (credentials, tokens, tool execution, and output handling), and consider isolation and least-privilege controls when experimenting with LLM-driven command execution in security tooling environments.
Jun 29, 2026
Two new **AI-assisted offensive security** efforts were highlighted: *BlacksmithAI*, an open-source penetration testing framework, and a **Kali Linux** workflow that enables natural-language control of common pentesting tools via Anthropic’s **Claude**. BlacksmithAI is described as a hierarchical, multi-agent system where an orchestrator decomposes assessment goals and delegates work to specialized agents aligned to typical pentest phases (recon, scanning/enumeration, vulnerability analysis, exploitation, and post-exploitation), then compiles results into reporting. Separately, Kali Linux was reported to support a native AI-assisted workflow by integrating Claude through the open-source **Model Context Protocol (MCP)**, using an `mcp-kali-server` package to translate plain-English prompts into tool execution (e.g., Nmap/Gobuster/Metasploit), parse outputs, and return structured findings and prioritized reporting. Both items reflect a broader shift toward **agentic and conversational interfaces** for offensive tooling, reducing reliance on manual CLI syntax while automating planning, execution, and summarization of assessment steps.
Jun 29, 2026
NVISO published a technical introduction on **automating LLM red teaming** to find security weaknesses in LLM-based applications, focusing on AI-specific risks such as **prompt injection**, **data leakage**, **jailbreaking**, and other behaviors that can bypass guardrails. The post describes why manual testing is difficult due to LLMs’ probabilistic behavior and demonstrates using the *promptfoo* CLI to scale testing against a deliberately vulnerable *ChainLit* application, positioning automated test harnesses as a way to systematically probe LLM apps for exploitable failure modes. Separately, a practitioner write-up describes how security analysts and engineers are using general-purpose LLM tools (*Claude*, *Cursor*, *ChatGPT*) to accelerate day-to-day security work through better prompting patterns rather than “keyword searching.” It provides practical prompting techniques (e.g., “role-stacking” and supplying richer context like requirements docs or code repositories) and includes an example of using an LLM to help design a small Flask application for collecting OSINT (DNS, WHOIS/RDAP, HTML) for URL investigations—guidance that is adjacent to, but not the same as, automated red-teaming of LLM applications.
Jun 29, 2026