CrowdStrike Identifies Three Tool Poisoning Attack Classes Targeting MCP AI Servers

·13 views

CrowdStrike has published a taxonomy of three distinct attack classes targeting Model Context Protocol (MCP) servers, the infrastructure layer that AI agents use to select and call tools. Attackers can embed malicious instructions inside tool description fields, manipulating AI agents into exfiltrating secrets or altering behavior without modifying any underlying code. Because the vulnerability exists in natural-language metadata rather than executable code, traditional static analysis and SAST tools fail to detect it. A second attack class, known as shadowing, allows one tool's description to corrupt how an agent constructs parameters for a completely separate tool later in the same session. CrowdStrike's Falcon Guardian, announced this month, is described as the first production-level tool capable of tracing prompts through tool calls to detect such prompt-layer threats.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Dropbox Joins Google Gemini as Connected App for File Search and Automation

Dropbox has launched as an official connected app for Google Gemini, allowing eligible users to find, summarize, share, and manage Dropbox files directly within Gemini Chat. The integration also extends to Gemini Spark, Google's multi-app workflow automation environment, where Dropbox can serve as the file layer alongside other third-party services like Canva and Instacart. The rollout is currently limited to US-based users who are 18 or older, use a personal Google account, work in English, and access Gemini via the web app at gemini.google.com. Access to Gemini Spark is further tied to Google AI Pro and AI Ultra subscription plans, meaning neither a Dropbox business account nor a work Google account alone qualifies a user. The integration does not replace Dropbox as a storage platform but aims to reduce context-switching by making Dropbox content accessible within AI-driven conversations and workflows.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Why Debugging Your Trading Backtest Matters More Than Upgrading Your Model

Algorithmic trading backtests can produce misleadingly strong results when they inadvertently use information that would not have been available at the time of a simulated trade decision. A common cause is joining datasets on date columns alone, which can allow post-decision data — such as earnings released minutes after market close — to influence earlier trade signals. Developers are advised to track multiple timestamps per data point, including when information was published, received, and processed, to accurately determine what a strategy could have known. Even filtering retrieved documents cannot fully eliminate look-ahead bias if an underlying model was trained on data that includes future outcomes. Rigorous validation of data eligibility, execution assumptions, and model complexity is essential before trusting any backtest result.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Combine Webhooks and Scheduled Polling to Minimize API Key Revocation Gaps

Engineering teams managing scoped API keys face a critical challenge: ensuring revocations propagate reliably to edge gateways that cache key status locally. Webhooks deliver revocation events within seconds but carry a hard risk — once a sender's retry budget is exhausted, dropped events leave no trace. Scheduled polling inverts this by giving the receiver control over liveness, though it introduces latency equal to roughly half the poll interval. Using both together means webhooks set the speed while periodic reconciliation sweeps verify what actually landed, with divergence counts serving as a live health signal for the push path. For audit purposes, the priority is not low median latency but the ability to reconstruct exactly when a key was revoked and how that revocation was confirmed — making an append-only intake log essential.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

LLM Basics Explained: Weights, Parameters, Tokens, and Sampling Controls

A Large Language Model (LLM) is fundamentally a collection of numerical weights derived from training data using architectures like Transformers, with model size measured in parameters — for example, a 120B parameter model contains 120 billion such weights. To run a model locally, sufficient RAM and CPU or GPU resources are required, with smaller models like 7B parameters being feasible on consumer laptops using tools such as Ollama or LM Studio. At its core, an LLM works as a next-word predictor, assigning probability scores to possible continuations and selecting outputs based on those scores. The randomness of output can be controlled through parameters like Temperature, Top-k, and Top-p, where lower temperature favors factual responses and higher temperature encourages creative or varied outputs. Text is broken into smaller units called tokens during a process called tokenization, and a model's context window defines the maximum number of tokens it can process at one time.

0 comments Read more at DEV Community