Open-Source Scanner Exposes Hidden Security Risks in AI Agent Tools

·3 views

A security scan of popular open-source TypeScript AI agents has revealed that many tools with innocuous names—such as grep, calculator, and query—actually execute shell commands, run eval, or trigger destructive HTTP actions. The tool, released as a no-install npm package called agentx-core/scan, analyzes an agent's codebase to inventory every model-accessible function and rank it by potential risk. Common findings included fetch tools vulnerable to server-side request forgery, email tools where the model controls recipients, and calculators that run arbitrary code via eval. The scanner reads actual function bodies rather than relying on tool names, and explicitly flags uncertainty rather than fabricating risk assessments. Developers are advised to run the scan and use the accompanying AgentX gateway to guard high-risk tools before deploying agents with expanded capabilities.

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)

Solo Dev Tackles AI Context Gap When Specs Live Across Multiple Repos

A developer building personal projects as separate microservices ran into a recurring problem: AI coding tools like Cursor can only access files within the currently open workspace, making cross-repo spec references difficult. Two broad approaches exist — adding all relevant repos to a multi-root workspace, or having the AI fetch files by path on demand — but each carries trade-offs in noise, token usage, or context bloat. Cursor's April 2026 update improved multi-root workspace support for cross-repo edits, yet the core issue of indexing entire repos just to read a few spec files remains. The developer also evaluated self-hosted documentation tools and external SaaS options, finding them either too rigid, too costly, or ill-suited for a solo workflow. Unable to find a satisfactory off-the-shelf solution, the developer began exploring building a lightweight custom tool to bridge the gap between isolated repos and AI context windows.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Engineer migrates live IoT smart cane platform to Google Cloud with zero forced app updates

In January 2024, a backend engineer took over WeWALK, a smart cane platform serving 27,000 visually impaired users, and migrated it from a largely inaccessible self-hosted Rancher/Kubernetes cluster to Google Cloud Run, Cloud SQL, and Firestore. The core challenge was ensuring no user was ever forced to update their app, since the cane is a mobility aid and any interruption could leave someone stranded. To preserve existing AWS WebSocket endpoints baked into older client binaries, the engineer built a Lambda-to-Cloud-Run proxy that silently forwarded traffic to the new infrastructure without clients noticing the change. The migration also involved reconciling heterogeneous data stores — MongoDB, PostgreSQL, and Firestore — while handling a continuous stream of 102-byte device diagnostics arriving every ten seconds from canes in the field. The entire cutover was completed with no reported support tickets related to the migration.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Four-Part Prompt Structure Found to Outperform Vague Prompts on Costly AI Models

Developers shipping AI features have identified a consistent pattern that distinguishes reliable prompts from unreliable ones. The framework requires every prompt to specify a role, a single task, its constraints, and an exact output format. This structured approach proved so effective that a precise prompt on a cheaper model regularly outperformed a vague prompt on a more expensive one. The finding challenges the assumption that model quality alone determines output reliability. The pattern has been validated across use cases including classification, extraction, chain-of-thought reasoning, and structured JSON generation.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Google's AI Overviews Cut Top-Ranked Site's Organic Traffic by 59%, Data Shows

A website holding the number-one Google ranking for 11 months saw its click-through rate plunge from 27% to 11% after Google rolled out AI Overviews broadly in the US market, resulting in a 59% traffic drop on affected queries. Analysis across multiple clients revealed a consistent pattern: queries where AI Overviews deliver a complete answer saw CTR fall by 60–80%, while queries requiring user action saw only a 10–25% decline. The findings expose a growing disconnect between search rankings and actual site traffic, as ranking tools continue to show top positions while clicks quietly erode. Notably, once AI Overview-affected queries were separated from others, the same client's remaining organic traffic had actually grown 12% in the same period. In response, the SEO team has begun shifting content strategy away from purely informational articles toward interactive tools and resources that give users a concrete reason to click through.

0 comments Read more at DEV Community