Agent-Based Scanner Catches 95% of Malicious AI Skills vs. 8% for Static Tools

·1 views

Researchers at Trent AI benchmarked five ClawHub skill scanners against a manually labelled set of 60 skills, covering benign, vulnerable, and malicious categories. Their agent-based scanner, trentclaw, achieved a 94.6% recall rate — over 40 percentage points higher than the next-best tool, ClawScan, which caught 54.1%. Signature and static analysis scanners performed worst, flagging as few as 8.1% of dangerous skills. The study found that reasoning-based scanners outperform pattern-matching tools partly because some malicious skills ship with no executable code, making them invisible to signature checkers. The benchmark comes amid real-world threats like the February 2026 ClawHavoc campaign, in which disguised productivity skills were used to steal API keys and browser credentials from users.

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)

Why Developers Should Understand One Layer Below Their Abstractions

Many developers learn tools like async/await, smart pointers, or high-level data types without understanding the underlying systems they abstract. This gap in foundational knowledge — spanning memory management, data types, and concurrency — can lead to unpredictable bugs and performance issues that are hard to diagnose. Abstractions work well most of the time, but they can obscure root causes when systems behave unexpectedly under pressure. The article argues that developers do not need to become hardware engineers, but should make a habit of understanding one level deeper than their current working layer. Learning through real problems, running focused experiments, and using tools that expose system behavior are recommended as practical ways to build this deeper intuition.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How to Build Reliable Login Recovery Systems Without Losing Account History

For e-commerce platforms, email and phone verification should be maintained as separate, auditable recovery factors rather than interchangeable channels, according to a developer guide on login risk scoring. The choice of verification channel should be driven by device-risk assessment, and a previously verified factor must not be silently overwritten by a new device fingerprint or network change. Every verification attempt should be treated like a financial transaction, assigned an idempotency key, an immutable audit record, and a clear expiry to handle duplicate sends and callbacks reliably. Before switching verification providers, teams are advised to export a detailed channel-level delivery ledger covering at least one full business cycle, broken down by country, carrier, device-risk band, and recovery outcome. Retaining only normalized, encrypted evidence needed to reproduce a decision — rather than raw contact data or full fingerprints — is recommended as a balanced approach between investigative utility and privacy compliance.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Kafka Streams Topology Builder Lets Developers Visualize and Run Pipelines Visually

A new approach to building Kafka Streams topologies allows developers to design processing pipelines on a visual canvas by dragging and wiring nodes such as source, mapValues, filter, and sink operators. The tool connects to a schema registry to automatically resolve Avro serializers and deserializers, enabling schema-aware field completions when writing transformation expressions. Developers can write mapValues logic using Spring SpEL, with completions drawn directly from the registered schema fields, while downstream nodes like filter can reference computed fields inferred from prior steps. Live records pulled from the actual topic flow through a preview, giving real-time feedback on how data transforms across the topology. The goal is to close the gap between sketching a stream-processing idea and having a fully runnable, wired topology without boilerplate build and serde configuration.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Zyte Meetup to Address Prompt Injection and Security Risks in Unattended AI Agents

Zyte is hosting a virtual Developer Community Meetup on September 24, 2026, in partnership with Humanbound, focused on securing AI agents that operate without human oversight. Unlike supervised agents, unattended agents that browse the web and execute tasks autonomously are vulnerable to prompt injection attacks, where malicious content embedded in fetched pages can manipulate the agent's behavior. Such attacks require no traditional exploit — an agent may be tricked into leaking sensitive data or escalating privileges simply by reading a compromised page or document. The session, titled 'Ship Agents That Survive the Real Web,' aims to demonstrate real-world agent failures and their fixes live. Attendees will learn how to map untrusted content entry points, build adversarial tests, and maintain reproducible agent definitions to improve security.

0 comments Read more at DEV Community