Browser Tool Tests LLM Security Using Deterministic Grading Without an AI Judge

·1 views

A developer has released The AI Crash Test, a browser-based tool that evaluates large language models against adversarial prompts using deterministic, code-based grading rather than a secondary AI judge. Users supply their own API key, which is sent directly to the model provider and never passes through the tool's server — a claim the developer says anyone can verify via browser DevTools in about 30 seconds. The grading engine, called gradecore, produces byte-identical scores across repeated runs, eliminating variability from temperature or judge drift. The tool covers eight tasks across seven attack categories, including prompt injection, hallucination baiting, and refusal calibration, generating a severity-weighted vulnerability report. The developer acknowledges that more comprehensive red-teaming tools such as NVIDIA's garak and Microsoft's PyRIT exist, positioning this project not as a new category but as a narrower, auditable alternative focused on reproducibility.

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)

AgentScaffold Adds Memory, Peer Review and Learning Loops to AI Coding Agents

A software engineer with backgrounds at Boeing, Salesforce, and Dropbox has released AgentScaffold, an open-source Python package designed to address three core weaknesses in AI coding agents. The tool tackles the lack of persistent memory across sessions, inconsistent planning discipline, and the inability to learn from past mistakes. AgentScaffold runs as an MCP server, making it compatible with tools like Cursor, Claude Code, and Windsurf, and uses DuckDB with a property graph to index codebases across eight programming languages. It also ingests governance artifacts such as plans, architecture decision records, and review findings, linking them directly to the relevant code. The framework aims to bring engineering-process discipline to AI-assisted development by ensuring mistakes are caught before shipping and standards improve over time.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Engineering Student Builds WireGuard VPN Server from Scratch on a Local VM

An engineering student undertook a self-directed project to build a WireGuard VPN server from scratch, aiming to learn real-world networking concepts beyond standard tutorials. Rather than using a cloud provider, the student ran a Debian virtual machine locally via VMware Workstation, which introduced challenges around virtual networking, NAT, and IP forwarding. Initial setup appeared simple but quickly revealed issues including disabled IP forwarding, misconfigured firewall rules, and incorrect routing that required deep troubleshooting. The student later rebuilt the VPN using Docker to compare containerized and native deployments, and tested connectivity across multiple environments including mobile hotspots and different operating systems. All findings, scripts, and configuration examples were compiled into a public GitHub repository to serve as a reference for others undertaking similar projects.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Five Major AI Models Launched in 16 Days, Signaling New Era of Parallel Competition

Between July 1 and July 16, 2026, five frontier AI models reached broad availability in rapid succession, with releases from Anthropic, xAI, OpenAI, Meta, and Moonshot AI all arriving within roughly two weeks. According to Artificial Analysis, six separate labs now field models scoring above 50 on the Intelligence Index, a sharp rise from just a handful of leaders months earlier. The burst of activity marks a shift away from the historical pattern of one lab holding a clear lead for months before rivals caught up. Anthropic's Claude Fable 5 return was notable for regulatory reasons, having been temporarily suspended over export-control concerns before being restored with additional safeguards — underscoring that frontier model deployments are increasingly subject to government oversight. Analysts note that as raw capability gaps narrow, factors such as cost, latency, and ecosystem integration are becoming the primary differentiators for developers and businesses choosing between models.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How a browser strategy game solves networking without UDP or prediction

A developer building Old Light, a multiplayer browser-based strategy game, has detailed the networking architecture designed for slow, long-running gameplay rather than real-time action. Unlike typical game networking that relies on UDP and client-side prediction, the game uses WebSockets with just two message types — a full world snapshot on connect and incremental delta patches thereafter. A key challenge was ensuring consistency at the boundary between connection and live updates, solved by buffering early deltas on the client and sequencing broadcasts carefully on the server. The developer also identified a subtle bug where resource counters double-counted income by pairing a freshly projected balance with a stale database timestamp, fixed by always anchoring projections to the moment they are calculated. The post highlights how slow-paced games face entirely different correctness problems compared to fast games, with bugs that can go undetected for hours rather than frames.

0 comments Read more at DEV Community