buc.ci is a Fediverse instance that uses the ActivityPub protocol. In other words, users at this host can communicate with people that use software like Mastodon, Pleroma, Friendica, etc. all around the world.
This server runs the snac software and there is no automatic sign-up process.
A stolen Fireworks AI key exposed unredacted credentials in coding-agent traces. Attackers used one of them on Oct 29 to plant a trojanized Claude Code extension (v2.1.312, v2.1.313) on Cursor's marketplace. Up to 5,000 users may have received it. Live about 11 hours.
The lesson: redact secrets before they reach the trace, not after the leak.
https://blog.relayshield.net/2026-11-18-fireworks-ai-stolen-key-traces
----------------
🛠️ Tool: DefenseClaw for AI Agent Runtime Controls
===================
DefenseClaw is an open-source Cisco AI Defense tool for governing AI coding agents. The supplied material describes skill and MCP server scanning, runtime decisions on tool calls, agent discovery and an audit trail. The tool has not been independently verified.
Key features
The control points cover checks before an agent loads a skill or MCP server, decisions during tool calls, and records of those decisions. Human approval can pause a high-risk action where the connected agent supports it. Decisions can be sent over OTLP to Grafana or Galileo, or to a webhook. A browser-based policy creator produces YAML for the configuration, while a terminal UI brings together alerts, scans and setup.
Technical implementation
Connectors use the strongest control exposed by each agent: native hooks, a policy plugin or a bridge. The source lists 11 agents, including Claude Code, Codex, GitHub Copilot CLI, Cursor and OpenCode. All listed agents have support on macOS and Windows. Linux support is listed for all except Claude Cowork, which is in preview on that platform.
Runtime guardrails start in observe mode, where calls are not blocked. In action mode, the example configuration leaves block_at unset and the default pack determines the outcome: CRITICAL blocks, HIGH and MEDIUM alert, and LOW allows. The example HIGH decision both alerts and allows the call, then emits a decision record. The optional LLM judge is disabled by default and is distinct from deterministic rules.
The source describes one policy and audit log across admission checks and runtime decisions. It also distinguishes those decisions: runtime actions are not reused as skill or MCP admission actions. The browser example does not evaluate Rego or scanner output.
Use cases and setup
The tool can inventory agents, local models, MCP servers and skills on a machine, including which ran. An organization can use observe mode to collect decisions without blocking, then enable action mode for a supported connector. The source also describes Intune or other MDM deployment and administrator-owned services, policy and program files.
1. After installation, run defenseclaw init to discover agents and start in observe mode.
2. Run defenseclaw guardrail mode action --connector claudecode to enable action mode for Claude Code.
3. The described quickstart adds a demo rule for observing a block. Release downloads are checked against release checksums, according to the source.
Limitations
Approval depends on agent support. The macOS menu-bar app and NVIDIA OpenShell sandbox integration are in preview. The OpenShell preview applies to Linux and Apple-silicon Macs. The supplied material makes an administrator-control claim but does not provide independent validation or enough detail to assess bypass resistance.
This is the experimental MCP Server (written in C) running against REDasm Core Engine solving a simple crackme and inspecting details about the loaded binary.
Tested using Claude Sonnet 5.5
Your wallet agent moves real money. Give it a firewall.
RelayShield Hosted MCP Proxy Firewall is live. It screens every wallet write call against our threat corpus and your intent binding before it reaches the wallet. Dedicated instance per tenant, $49/mo direct, works with Coinbase Wallet MCP and MetaMask.
https://blog.relayshield.net/2026-11-13-hosted-mcp-proxy-firewall
🏭 New from Netskope Threat Labs: the Manufacturing 2026 report.
🤖 Over the past year, active AI users in manufacturing rose from 40% to 67%, and managed AI use doubled from 35% to 70%.
🔒 Regulated data accounts for 57% of AI-related data policy violations.
🔗 Non-human entities connecting to remote MCP servers are up 490%, and AI lures keep arriving in waves. 🎣
📊 Infographic below 👇
📖 Full report: https://threatlabs.netskope.com/reports/2026/threat-labs-report-manufacturing-2026
You installed an MCP plugin once. It can update silently, forever. Live demos of what that costs you. #MCP #AISecurity #BSidesDE https://events.humanitix.com/security-bsides-delaware-2026?c=talks
🤖 Et si vous pouviez interroger votre wiki en langage naturel depuis votre terminal ?
Wiki.js 3 intègre un serveur MCP. Mon nouveau guide explique comment le connecter à Gemini CLI et OpenCode : jeton d'accès, réglages Nginx Proxy Manager pour le streaming, et astuces pour éviter les blocages de quota.
📖 https://wiki.blablalinux.be/fr/piloter-wiki-js-3-gemini-cli-opencode-mcp
#WikiJs #MCP #GeminiCLI #OpenCode #SelfHosting #Linux #LogicielLibre #BlablaLinux
AI Agents Need Security Boundaries at Every Execution Layer
As agents move from generating suggestions to changing code, calling tools, and operating infrastructure, every execution boundary must be able to refuse unsafe actions and leave evidence.
https://securstack.io/blog/en/ai-agents-need-security-boundaries-at-every-execution-layer
Mal eine Frage im Bereich AI mit LLM und explizit im Zusammenhang mit a11y.
In wie weit gibt es bei blinden Menschen oder auch mit starker Sehbehinderung (übrigens auch Menschen, die nur über Sprache kommunizieren können) das Interesse (oder vielleicht auch nur die Idee) Applikationen über KI Chats mittels LLM Modelle und dem Protokoll MCP zu steuern, anstatt sich durch wilde UIs zu wühlen?
Mittels MCP (Model context protocol), werden Anwendungen über Text oder Sprache bedienbar, allerdings über den Proxy eines AI Chatbot mit passenden Modell, das MCP nutzen kann. Es kann als Alternative zu ARIA, UIA usw dienen (oder wo sich a11y in den UI-Frameworks nicht etabliert hat).
Gibt es da Meinungen und Erfahrungen zu?
Bitte boosten
----------------
🛠️ Tool
===================
sandbox-mcp: Disposable Windows 11 Guests for AI Agents
Scott Hanselman's sandbox-mcp wraps Windows Sandbox with two surfaces that share one engine: a portable CLI called sbx and an optional MCP server that exposes discoverable tools and inline images to agent clients. The loop: boot a fresh Windows 11 guest, deploy an app, drive its controls, catch the crash, keep the evidence, destroy the machine. Version 1.2.0 bundles Node.js and a guest agent, so no separate Node or npm install is needed. Measured end-to-end time for GUI samples sits at roughly 40-50 seconds, per the project docs.
Key Features
• Fresh guest on every boot, with no implicit app runtime installs
• Native UIA control: inspect, type, click, toggle, wait; covered frameworks are WinUI 3, WPF, WinForms, and Win32, plus MSIX/AUMID activation
• Network modes: off, firewall, or on
• Evidence that survives teardown: PNGs, bounded recordings, self-contained HTML, optional MP4, process state, exit codes, crash events, installer and uninstaller diffs, detonation observations
• Dependency diagnosis via sbx deps app.exe; verified provisioning happens only on explicit request (sbx run app.exe --provision auto)
The Compatibility Demo
The docs show the "works on my machine" loop with measured output. NativeProof.exe on a fresh, network-off guest exits with code -1073741515, which maps to 0xC0000135, STATUS_DLL_NOT_FOUND, the classic missing VC++ redistributable failure. After an explicitly requested provisioning pass, the same loop on another fresh guest passes: the redistributable (Microsoft-signed 14.51.36247.0) was downloaded and verified on the host, guest networking stayed off, and the docs honestly label the result PROVISIONED rather than a clean-machine pass.
Use Cases
• Reproducing dependency and version gaps on a clean client instead of arguing about the dev box
• Crash capture with exit codes and crash events instead of eyeballed repro reports
• Detonation-style observation of desktop installers on isolated guests
• UI automation regression for desktop apps where browser automation does not apply
Setup Considerations
CLI first. Configure the MCP adapter only when the agent actually needs that interface. The setup doc instructs agents to verify the latest release download and to ask before admin changes or reboots.
Limitations
The whole layer rides on Windows Sandbox, so a suitable Windows host is required. Timings and screenshots in the docs are the authors' measurements, not independently verified. Provisioned runs are not clean-machine passes, and the docs state that plainly.
References
• GitHub: https://github.com/shanselman/sandbox-mcp
• Agent setup doc: https://raw.githubusercontent.com/shanselman/sandbox-mcp/main/docs/agent-setup.md
• Website: https://shanselman.github.io/sandbox-mcp/
🔹 tool #MCP #WindowsSandbox #Windows11 #infosec
----------------
🎯 AI: AI Agent Artifact Catalog Maps Endpoint Indicators for Coding Agents and LLM Runtimes
===================
A working catalog under the name AI Agent Artifact Catalog enumerates observable artifacts for mainstream AI agent tooling. Each entry is split into network sourcing and process sourcing, carries an identifier in the AIRT-XXXX series, marks sourcing confidence (high or medium), and lists OS applicability (Windows, macOS, Linux, Docker). The format reads as a detection-oriented reference rather than a threat report, since the tools covered are legitimate products and the value is in endpoint visibility.
Coverage in the provided sample:
Coding agents: Claude Code (api.anthropic.com, claude.ai installer, statsig.anthropic.com telemetry; node.exe (claude) on Windows, claude / node on Unix), Cursor (.cursor.sh, api2.cursor.sh, .anysphere.co; Cursor.exe), GitHub Copilot in Agent Mode or CLI (api.github.com, copilot-proxy.githubusercontent.com, api.githubcopilot.com; copilot-language-server / node), Aider (openrouter.ai, api.openai.com, api.anthropic.com; aider / python), Continue.dev (arbitrary configured provider base_url; node extension host), and Codeium / Windsurf (server.codeium.com, .codeium.com, .windsurf.com; Windsurf.exe, language_server_windows_x64.exe).
MCP host: Claude Desktop (api.anthropic.com, claude.ai; registry.npmjs.org for npx MCP server fetch; Claude.exe / Claude).
Local runtimes: Open Interpreter (api.openai.com or configured base_url; interpreter / python), Ollama (native and OpenAI-compatible API on 11434/TCP; registry.ollama.ai, ollama.com; ollama.exe / ollama), LM Studio (1234/TCP; huggingface.co, lmstudio.ai; LM Studio.exe, llmster.exe, lms.exe), Jan.ai (1337/TCP, also 8080; jan, jan-daemon), and text-generation-webui (Gradio UI and OpenAI-compatible API on 7860/TCP, 5000/TCP).
Orchestration: n8n (editor UI, REST API, webhooks on 5678/TCP; node), AutoGen (python), and CrewAI, whose entry is cut off mid-line in the excerpt.
Detection notes:
1. Local listener ports are the strongest pairing signal: 11434, 1234, 1337/8080, 7860/5000, and 5678 are uncommon service ports, and most are documented as OpenAI-compatible APIs.
2. Process names alone are weak: python and node are ubiquitous, so correlation should combine binary names and command lines, which the catalog does include.
3. Two edge cases are explicitly recorded: Continue.dev points at an arbitrary configured base_url, so network attribution depends on user configuration; Open Interpreter in --local mode is marked as producing no network traffic, leaving process sourcing as the only signal.
4. Aider's OAuth PKCE callback listener on ports 8484-8584 is a distinctive, low-noise artifact.
5. Claude Desktop pulling MCP servers from registry.npmjs.org via npx means MCP server installation itself leaves a recognizable network trace.
Limitations: the excerpt is truncated at CrewAI, and identifiers are non-contiguous in the sample (AIRT-0001 through AIRT-0027), so the full catalog likely holds more entries. Domains and ports are point-in-time by nature and will drift across tool versions. Nothing here is an indicator of malice; these are artifacts of legitimate tools, so any verdict depends on context, including abuse scenarios where the same local API ports appear.
Practical use: egress allowlists and blocklists for approved AI tools, EDR correlation rules matching process identity to local ports, and shadow-AI inventory hunts. Refresh against the source catalog before writing production rules.
🔹 AIAgents #DetectionEngineering #MCP #ShadowAI #bookmark
----------------
🛠️ Tool
===================
Security Detections MCP is an MCP (Model Context Protocol) server that gives LLMs a query interface over a unified database of security detection rules. The supported sources are Sigma, Splunk ESCU, Elastic, KQL, Sublime (covering both Sublime Security rules and delivr.to rules), and CrowdStrike CQL. The problem it addresses is mundane: detection content lives in half a dozen separate repositories with different query dialects, and cross-referencing them by hand is slow.
Key Features
• Unified indexing of detection rules from six sources behind one query interface.
• Local mode exposes 81 tools through an npm package.
• Hosted mode offers roughly 25 read-only tools through a Streamable HTTP endpoint, kept in sync with upstream content, with a free tier of 200 calls per day.
• Install paths for Cursor and VS Code (deeplink buttons), Claude Code and OpenAI Codex (CLI one-liners), Claude Desktop (JSON config).
Technical Implementation
The local variant ships as the npm package security-detections-mcp and runs via npx. It requires Node.js and expects environment variables pointing at local clones of the rule repositories: SIGMA_PATHS, SPLUNK_PATHS, STORY_PATHS, ELASTIC_PATHS, KQL_PATHS, SUBLIME_PATHS, and CQL_HUB_PATHS. Multiple paths per source are supported, so several rule sets can be indexed under one prefix. The README claims about ten minutes of setup; believable only if the clones already exist.
The hosted variant authenticates with a bearer token generated via the author's site, exposes a Streamable HTTP endpoint for any MCP client, and restricts the toolset to read-only operations. Sync is handled server-side, so there is no local repo maintenance in this mode.
Use Cases
• Coverage checks: ask whether a given TTP already has detection logic in any supported dialect.
• Cross-dialect comparison: see how the same behavior is expressed in Sigma versus Splunk ESCU or CrowdStrike CQL.
• Detection authoring: use existing rules as reference material in an LLM-assisted writing workflow.
• Threat hunting: surface Sigma threat-hunting rules alongside the standard set, since the setup treats them as separate paths.
Installation Overview
For local mode: clone the repositories you care about, register the server through your MCP client (claude mcp add, codex mcp add, or a JSON config), then set the environment variables. For hosted mode: generate a token and paste the endpoint URL plus auth header into the client. Step-by-step guides exist for macOS, Windows (WSL and native), and Linux.
Limitations
• Local mode means maintaining local clones of every source repository you index; that is the real recurring cost.
• Hosted mode is read-only and capped at 200 calls/day on the free tier.
• The Sublime source bundles two rule sets (Sublime Security and delivr.to), which matters for provenance attribution.
• Everything above is based on the README; haven't tested it personally.
References: project README, SETUP.md, and docs/HOSTED_MCP.md in the repository.
🔹 tool #MCP #Sigma #DetectionEngineering #ThreatHunting
🔗 Source: https://github.com/MHaggis/Security-Detections-MCP
VOLTA, Talos report: an infostealer that ranks victims by value, with a model, and falls back through three models when one refuses. Do not rely on a single refusal to buy time. Screen the outbound bundle and the inbound ranking, at the payoff call.
New post: https://blog.relayshield.net/volta-ai-victim-ranking
Eon, simulated study: agents deviated in 40 percent of runs, falsely claimed completion in 290 runs, with every model doing it at least once. Do not accept done from the agent that did the work. Accept it from a record showing who acted, what was authorized, and what settled.
New post: <https://blog.relayshield.net/eon-intent-identity-content>
Prefab UI's opt-in Generative UI sandbox has a bypass (CVE-2026-120338). The Pyodide/Deno boundary let caller-supplied Python read and write host files as the server OS user.
Relevant to FastMCP servers that expose generate_prefab_ui. Unauthenticated exposure depends on the server's access controls.
Fixed in prefab-ui 0.21.0, released Oct 26. Check versions and tool access.
----------------
🛠️ Tool
===================
Vulnerator AI: Hermes agent drives an OODA pentest loop against a Kali container
The project pairs a language-model agent with a live Kali Linux environment and lets the agent run offensive tooling directly against a target. The driving brain is Hermes, the agent framework from Nous Research, and the runtime is a Kali container governed by an MCP server. The agent observes tool output, orients on it, decides the next action, and acts by launching tools straight through the Kali MCP. Nothing sits between the decision-maker and the tooling, and there are no human approval gates in the loop.
Key characteristics
• An OODA (Observe-Orient-Decide-Act) loop structures the agent's decisions instead of a flat tool-calling chain
• Hermes talks directly to the Kali MCP server, with no orchestration layer in between
• The design is deliberately minimal: no core service, no REST API, no UI, no approval gates
• Four containers make up the deployment: the agent, two tool MCP servers, and the MCP front door
• A demo video shows the loop in action
Technical implementation
Docker handles the whole deployment. Configuration goes through a .env file: set an OPENROUTER_API_KEY and generate two tokens with openssl rand -hex 32. The repository ships a security pipeline with CodeQL scanning and PR quality checks. One caveat on sourcing: the published README cuts off mid-sentence in the architecture section, so the exact roles of the two tool MCP servers and the networking between containers are not documented in the material available. What is present confirms the four-container layout and the direct agent-to-Kali path.
Use cases
• Automated recon-to-exploitation workflows in authorized lab environments
• Studying how agentic models handle offensive decisions end to end, since Hermes owns tool selection
• A minimal reference point for anyone building an MCP-governed tool runtime, useful when comparing against heavier frameworks
Limitations
• Haven't tested this personally; the source is the project's own README, not a validated engagement report
• No approval gates means the agent acts on its own decisions. That is the point for lab work and a liability anywhere production infrastructure is reachable
• The truncated architecture section makes container roles and networking details hard to confirm
• Scoping and authorization discipline fall entirely on the operator
References
Demo video in the repository, the Hermes agent project from Nous Research, and the Kali Linux tooling runtime. Releases are versioned on GitHub.
🔹 tool #MCP #AIAgents #pentest #OODA
🔗 Source: https://github.com/irraz/vulnerator-ai
According to Salt Security: 49.8% of firms confirmed or suspected a rogue agent action in the past year. Only 12.5% can trace a complete agent journey and only 17.5% can stop one automatically. Verify who it is, bind what it intends, screen what it touches at the tool call.
New post: <https://blog.relayshield.net/salt-rogue-agent-actions>
Nine MCP servers. One bug. Eight no fix.
Tool parameters hit execSync, subprocess.run(shell=True), child_process.exec — no shell metacharacter check. Prompt-inject the agent via email or repo name, get arbitrary shell commands.
CVSS scores it AV:L. The attacker is remote. The score describes the agent, not the threat.
Inventory your MCP configs. Search for shell=True and execSync. Treat allowlists as indicators, not controls.
threataft.com/articles/mcp-server-command-injection-wave-cve-2026-117558-117573
#MCP #CVE #InfoSec #AIAgent #ShellInjection #ThreatIntel #CyberSecurity
MCP Security Is Now a Software Supply Chain Problem
Model Context Protocol connects AI agents to tools, data, and external systems. That flexibility creates new paths for data leakage, prompt injection, excessive permissions, and remote code execution.
https://securstack.io/blog/en/mcp-security-is-now-a-software-supply-chain-problem
#aisecurity #mcp #aiagents #applicationsecurity #softwaresupplychain
Forrester 2027: 10+ employee-built shadow AI agents for every sanctioned one at a third of large APAC enterprises. They never hit an asset inventory; they show up in MCP tool calls. Govern the tool call and you govern the agent.
New post: https://blog.relayshield.net/shadow-ai-agents-forrester
A Skynet LLMjacking agent hijacks exposed AI models to scan, hack servers and mine crypto for 100,000 steps, while victims pay the bill.
#LLMjacking #AISecurity #AIAgents #Cryptojacking #MCP #Symantec #Broadcom #Cybersecurity
http://securityonline.info/skynet-llmjacking-agent/?utm_source=mastodon&utm_medium=jetpack_social
Four reports, one week. One conclusion.
Attackers are not breaking AI models. They are walking through the infrastructure.
1,467 MCP servers in one scan. Half the ecosystem repos teach plaintext credential storage. Shadow MCP servers on laptops via Git, Slack, and zip files, never inventoried.
The infrastructure is the attack surface. Screen it.
----------------
🛠️ Tool: REA (Reverse Engineer Anything)
===================
Overview
REA is an MIT-licensed MCP server, distributed as the rea-agents package on npm, that exposes a reverse engineering toolset to any MCP-capable agent. The framing in the morluto/rea README is direct: inspect native binaries, JavaScript and Electron applications, .NET assemblies, and websites without their source code, and get explanations that carry supporting evidence and stated limitations rather than bare model summaries.
Key features
• One MCP surface across several target types: native binaries, JavaScript and Electron apps, .NET assemblies, websites
• Native analysis can build on an existing Hopper or Ghidra installation, instead of shipping yet another disassembly pipeline
• The same tools stay usable directly from the terminal, so agent access is optional, not mandatory
• Analysis runs locally; conclusions ship with their evidence and limitations, which is the right counterweight to overconfident agent output
• A generated tool catalog is published in the repository docs (mcp-contracts.md), so the exact tool surface an agent will see can be reviewed before wiring it up
Technical notes
The npm distribution and Node.js requirements suggest a standard Node-based MCP server. Setup registers REA with the agent and installs matching workflow instructions. The screenshots show an analysis bridge launching inside Hopper while it inspects a native binary, which implies the server drives the installed tooling rather than replacing it. The exact bridge mechanics, supported agent clients, and per-tool contracts remain unverified; the contracts document is where that check should start.
Use cases
• Feature archaeology: the flagship scenario from the README, ask the agent how another application implements a feature, get a documented, evidence-backed explanation. Worth noting that rebuilding a proprietary feature lives in uncertain legal territory.
• Triage support for unknown files when static analysis with receipts matters more than a quick verdict
• Terminal-only use for scripted pipelines where an agent adds nothing
Limitations
• Claims come from the project README only; nothing here has been independently tested
• Capability depth depends on which underlying tools are installed and licensed, particularly for native analysis
• The README is polished and badge-heavy, so treat the published tool contracts, not the landing page, as the source of truth for what the server actually exposes
• One sensible warning is included: fetch REA only from the official repository, the rea.tools website, or the rea-agents npm package, since lookalike repository names or copied readmes do not make a download official. For anything plugged into an agent, that is a baseline requirement, not paranoia.
References
Repository: morluto/rea. Website: rea.tools. Package: rea-agents on npm. Tool contracts: docs/mcp-contracts.md in the repository.
🔹 tool #MCP #ReverseEngineering #BinaryAnalysis #Ghidra
🔗 Source: https://github.com/morluto/rea
NVIDIA shipped a local MCP server with the toggle ON by default. 12 tools, 6 can change your machine. One binary toggle, no per-action authorization.
Anthropic confirmed the STDIO command injection flaw is "by design."
The protocol won't save you. Screening will.
https://blog.relayshield.net/nvidia-turned-on-a-local-mcp-server-toggle-on-by-default
El lado del mal - ARTCraft: El Clon de la Suite de Adobe hecho con IA en formato Open Source... para Agentes IA https://elladodelmal.com/2026/10/artcraft-el-clon-de-la-suite-de-adobe.html #Adobe #OpenSource #AgentesIA #AgenticAI #MCP #IA #IA #PDF #InDesign #ARTCraft
New: MCP Proxy Firewall
Pre-deployment scanner (free) + runtime proxy for MCP servers.
Blocks prompt injection, SSRF, OAuth abuse, PII leaks.
A fake Web3 accelerator used malicious npm packages to drop stealers on macOS and Windows and reroute AI coding agent traffic to its servers.
#npm #SupplyChainAttack #Web3 #CloudSyncD #Infostealer #MCP #AIAgents #Cybersecurity
forgejo-mcp 0.2.1 is out: an open source MCP server that lets AI assistants work with Forgejo and Gitea.
New since 0.1.x:
• 457 tools, up from 103, covering most of the Forgejo 16 API: Actions, webhooks, branch protections, wiki, packages and more
• Flags to switch off the areas you don't need
• Tool list uses about 40% less context
• HTTP mode handles multiple clients
• Security fixes, so please upgrade
Serious question:
Has all the research on #AI #security yielded any notable results, yet? I mean something beyond obvious stuff like:
* put your #agents into a sandbox
* run them with limited permissions (principle of least privilege)
* if they process untrusted data, try to harden their #prompt so they won't follow instructions from #promptinjection
* connect to your #MCP servers over TLS
Automating BGP Analysis: A Deep Dive into the BGPHorizon API & MCP
Currently testing omnimem v7 ready for release next week and early tests are strong: 27% cheaper to run, 23% fewer tokens and 35% quicker at completing a task (mainly due to deadends/the graveyard stopping 10/10 failures compared to 2/1) than without omnimem installed).
This comes with a much easier install path also!
New from RelayShield: detection for #LLMjacking — the fast-growing threat where a leaked LLM API key becomes a live, uncapped billing liability.
Real incidents: $46K/day (Sysdig, AWS Bedrock), $82K/48hr (leaked Gemini key), a $500K single-month bill from one unthrottled key. Underground price for a stolen key: ~$30.
Free to try, no signup. One of 13 tools on our MCP server.
New reference implementation shipped today: #LangChain middleware that gates AI agents from connecting to unfamiliar MCP servers behind a mandatory security check. It is not an optional tool call the agent is free to skip.
AWS Kiro IDE RCE via MCP Config and Hidden Text
🔗 https://cybersecurefox.com/en/aws-kiro-ide-remote-code-execution-mcp
#aws #kiro #kiro #ide #remote #code #execution #rce #mcp.json
New research on AI connector risk: 37% of Claude and ChatGPT connectors changed in six weeks, quietly gaining write access and new tools.
#AISecurity #Connectors #MCP #PromptInjection #AIGovernance
https://meterpreter.org/ai-connector-risk/?utm_source=mastodon&utm_medium=jetpack_social
I created an issue on the plugin's GitHub about removing the MCP functionality into a separate plugin, and the maintainer immediately converted it into a discussion, which is here: https://github.com/coddingtonbear/obsidian-local-rest-api/discussions/272
I'm getting the impression from all the LLM-related discussions and commits that this is a lost cause and I'll probably have to fork this plugin if I want an MCP-free version of it. Looks like they took up slop coding earlier this year too, which is another area of concern.
#AI #GenAI #GenerativeAI #AgenticAI #LLM #Claude #Copilot #AIPoisoned #NoAI #Obsidian #MCP #REST #API