Threat Library

The elasticsearch_memory tool in strands-agents-tools exposed connection parameters like the target host directly to LLM control, allowing a crafted prompt to redirect the tool to an attacker-controlled server. When the api_key parameter was omitted, the tool silently fell back to the operator's environment-stored Elasticsearch API key and sent it to whatever host the LLM specified, leaking the credential via the Authorization header. This is a classic tool-schema over-permissioning issue that turns an LLM-controllable field into a credential exfiltration vector.

The deprecated WebSocket transport in the MCP Python SDK accepted connections without validating Host or Origin headers, meaning any malicious webpage a victim's browser visits could open a WebSocket connection to a locally or network-exposed MCP server. This is a classic Cross-Site WebSocket Hijacking (CSWSH) pattern that could let an attacker-controlled origin interact with an MCP server's tools on behalf of an unwitting user. Severity is moderated because the affected transport is deprecated and impact depends on what the exposed server can do and whether it's reachable from a browser context.

The MCP Python SDK's experimental task management feature failed to bind tasks to the session that created them, allowing any connected client to list, read, cancel, or consume messages for tasks belonging to other clients. This is a broken access control / IDOR-style flaw that breaks the trust boundary between concurrent MCP sessions on the same server. It is fixed in version 1.27.2.

The MCP Python SDK's SSE and stateful Streamable HTTP transports route messages to sessions based solely on a session ID, without checking that the requesting client is the same authenticated principal who created that session. Any bearer-token-authenticated client that learns or guesses a valid session ID can inject JSON-RPC messages into another user's active session, effectively hijacking it. This is a serious cross-tenant authorization flaw fixed in version 1.27.2.

A researcher discovered a loophole in Anthropic's Claude web_fetch tool that allowed a malicious website to exfiltrate private user data (name, location, employer) by chaining together a sequence of attacker-controlled links discovered within previously fetched pages. This bypassed the intended safeguard that web_fetch could only follow user- or search-originated URLs, effectively encoding stolen data letter-by-letter into a series of outbound requests. Anthropic has since patched the issue by disallowing navigation to links found within fetched content.

Penpot's MCP server component exposed an unauthenticated HTTP endpoint bound to all network interfaces that allowed arbitrary JavaScript execution on the host. Any network-adjacent attacker could remotely execute code without credentials, making this a critical, easily exploitable vulnerability in an agent-tooling component.

Cursor's browser-enabled Cloud Agent exposed a local agent control endpoint without authentication, allowing malicious web content encountered by the agent's browsing capability to reach it from inside the sandbox. This let attacker-controlled pages trigger code execution within the agent's session and steal repository files, environment variables, credentials, and GitHub App tokens. The vendor fixed the issue by requiring authentication on the endpoint.

This is a defensive research paper, not a disclosed vulnerability or active exploit. It proposes PVDetector, a training-free method to detect prompt injection attacks against purpose-specific LLM agents by analyzing hidden activation states for 'policy-violation' concepts rather than just input/output text. Since the raw data describes a detection technique intended to help defenders, it does not itself constitute a threat.

PraisonAI versions before 4.6.78 default to running the MCP HTTP-stream server without any API key or authentication, meaning anyone who can reach the endpoint can list and invoke all exposed tools. The server also fails to validate tool-call arguments against the advertised schema, compounding the risk of malformed or malicious inputs reaching tool handlers. Exploitation requires the operator to have bound the server to a network-accessible address rather than the safe localhost default.

An unauthenticated attacker can abuse the Grafana MCP Server by injecting a crafted X-Grafana-URL header, tricking the server into acting as a proxy that leaks its own privileged Grafana service-account token. This same flaw allows server-side request forgery against internal networks and cloud metadata endpoints, giving attackers a path to full credential theft and internal reconnaissance without needing any prior authentication.

This is a Simon Willison blog post describing a fun, benign feature of OpenAI's Codex Desktop app: an animated desktop 'pet' (a pelican on a bicycle) created using gpt-image-2 and open-source skill scripts. There is no security vulnerability, exploit, or malicious activity described in this content.

The mastergo-magic-mcp MCP server (versions up to 0.2.0) contains a server-side request forgery vulnerability in its getComponentLink tool, caused by insufficient validation of the url argument passed to a z.string schema check. A remote attacker can supply a crafted URL to make the MCP server issue requests to arbitrary internal or external endpoints. A public exploit exists and the vendor has not yet responded to the disclosure.

The AWS HealthLake MCP Server fails to validate that pagination URLs returned via the next_token parameter point back to the legitimate HealthLake endpoint, enabling an authenticated user to redirect the server's outbound requests to an attacker-controlled endpoint. This can leak AWS temporary security credentials used by the MCP server, giving an attacker a foothold to access AWS resources tied to those credentials. Fixed in version 0.0.14.

This is a quoted essay excerpt from Armin Ronacher discussing how shared understanding in software projects is built through human friction, and how AI coding agents affect that process. It contains no technical details about vulnerabilities, exploits, or attacks against AI agents, and is purely a philosophical/editorial commentary piece.

Researchers demonstrate that malicious text placed in the physical environment (e.g., signs, labels) can be captured by camera-equipped smart glasses and hijack the behavior of Vision-Language Models, causing them to ignore true visual context and produce harmful, biased, or false outputs. This is a research paper describing a demonstrated but not actively exploited class of attack, with success rates up to 96% in simulation and 60% in real-world tests across 12 VLM models.

This is an academic benchmark paper (NetInjectBench), not an active exploit, demonstrating that LLM agents used for network operations can be manipulated via indirect prompt injection embedded in tickets, alerts, logs, and ChatOps messages to trigger unsafe tool actions. The researchers show naive agents execute unsafe actions 82.5% of the time under attack, but a metadata-aware execution-time policy gate can reduce this to near-zero while preserving usefulness. The severity is rated medium because this is defensive research quantifying and mitigating a known risk class rather than a disclosed vulnerability in a specific deployed product.

Researchers demonstrate that an adversary can poison an open dataset with misleading metadata and upload it to a public repository, causing autonomous AI research agents (built on Claude, GPT, Gemini) to unknowingly retrieve and use the poisoned data, producing fraudulent scientific conclusions in nearly half of tested runs. No prompt injection, agent compromise, or fabricated papers are needed — only manipulation of the open data ecosystem — and current agents rarely detect the poisoning (6% detection rate), though provenance auditing fully mitigates it in testing.

This is an academic research paper, not an active exploit or attack. The authors built MCPZoo, a large dataset of runnable MCP servers, and found that existing MCP security scanners are unreliable: they flag ~97% of servers as risky, but fewer than half of sampled alerts are true positives, and different scanners disagree substantially with each other.

This is a Simon Willison blog post sharing a GitHub code-frequency chart to illustrate how much AI coding agents (e.g. Opus 4.8, GPT-5.5) have increased his output on the Datasette open source project. There is no security vulnerability, attack technique, or agent threat described in this content.

MCP Appium, an MCP server for automating mobile app testing, fails to sanitize attacker-controlled UI element attributes before embedding them in an HTML template returned by its generate_locators tool. A malicious mobile app under test can inject HTML/JavaScript that executes in the MCP client's rendering context and calls arbitrary MCP tools via postMessage, effectively letting the app-under-test hijack the testing agent's capabilities.