Web Proxy TrickyBird Blocks WebMCP Tool Access Before Registering Its Own

·1 views

TrickyBird, a web proxy, had to address a security gap in WebMCP before implementing the browser-agent tool feature. Because all frames on a proxied page share the same origin, the default Permissions Policy would have allowed any ad frame to register tools into a user's session. The team fixed this by sending a tools=() Permissions-Policy header on every served document and removing document.modelContext from proxied pages via an injected shim. After locking down third-party access, TrickyBird registered a single tool called open_site on its own homepage, which navigates users to a proxied page when given a valid address. The tool is verified in Chrome 152 and scores correctly in Lighthouse 13.4.1, though real-world usage remains minimal as no known agent traffic has appeared yet.

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)

Developer Merges Creative and Technical Identities Under One Unified Vision

A software engineer and musician has publicly reflected on years of treating technology and creativity as separate pursuits, concluding the division was fundamentally flawed. The writer, who runs both the technical venture CitrusWorx and the music-focused Blackwater Sound, now frames artistic intent as the destination and engineering as the means to get there. This shift in mindset has reshaped how they approach multiple projects, including audio software, guitar building, and amplifier design under the Blackwater Sound umbrella. They argue that music and technology have always been inseparable, pointing to the long history of musicians repurposing and reinventing tools to achieve new sounds. The realization has also renewed their motivation to develop Kiwi Engine, a personal web application engine, now with a clearer creative purpose driving it.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Elementor Pro RCE Flaw CVE-2026-32475 Allows Unauthenticated File Upload Attacks

A critical unauthenticated remote code execution vulnerability, CVE-2026-32475, has been discovered in Elementor Pro, a widely used WordPress page-builder plugin. The flaw stems from a logic error in the file upload validation function, where a premature return statement skips extension and MIME-type checks for subsequent files in the same upload field, while a separate file-mover loop still processes and writes them. Any site hosting a published page with a non-required Elementor Form file upload field is exposed, requiring no authentication or nonce bypass since the form handler is accessible to unauthenticated users by design. The vulnerability was independently reported by Tin Pham via Patchstack on July 16 and by Austin Ginder via Wordfence's bug bounty program around July 24, with active exploitation observed in the wild during an attack peak between August 19 and 23. Site owners are urged to update immediately to Elementor Pro version 4.2.2 or later, and to add server-level rules blocking PHP execution under the forms upload directory as a defense-in-depth measure.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Amazon Bedrock Managed KB Matches Manual RAG Setup but Agentic Planner Falls Short

A developer benchmarked Amazon Bedrock's Managed Knowledge Base against a manually configured RAG pipeline built in April 2025, testing five retrieval configurations across single-hop and multi-hop question sets. The managed service's simple retrieval mode nearly matched the manual setup in correctness scores (0.88 vs. 0.84), and its Smart Parsing feature successfully ingested PDFs that had previously failed. However, the AgenticRetrieveStream planner using AWS's default model failed to decompose any query across all 80 evaluated questions, scoring lower than simple retrieval on multi-hop tasks. Switching to a custom large-model planner (Claude Sonnet 4.6) pushed multi-hop correctness from 0.50 to 1.00, though the author notes the planner and generator shared the same model in that test. The findings suggest managed retrieval is a viable alternative to manual chunking, but the agentic component's effectiveness depends heavily on which model handles the planning.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Why Recording Failed AI Tests as Zero Can Skew Your Brand Visibility Data

A data quality checklist published on DEV Community warns that AI visibility dashboards can misrepresent brand performance when failed or incomplete test sessions are logged as zero mentions rather than missing data. The guide distinguishes between three outcomes: a confirmed mention, a confirmed non-mention, and an unassessable result, arguing each must be stored and reported separately. It recommends tracking both a mention rate among completed answers and a collection coverage rate to avoid hiding gaps in the dataset. The checklist also cautions against re-running valid tests simply because a brand was absent, noting that cherry-picking favorable responses distorts what the metric actually measures. Additional edge cases flagged include negative brand comparisons, similarly named businesses, and plain-text domain references that were not used as cited sources.

0 comments Read more at DEV Community