SShortSingh.
TopTechnologyAI & MLStartupsProgrammingScienceBusinessFinanceCrypto & Web3World
Back to feed
Programming

Solo Dev Scraps Failed Multi-LLM Idea, Ships Lightweight Security Kernel for AI Apps

0
·1 views

A solo developer spent six months testing a 'mitosis' approach that split tasks across multiple LLMs, only to find it reduced test correctness from 95% to 83% at four to six times the cost across three independent experiments. Rather than ship the flawed feature, the developer deleted it and documented the failure publicly in a FINDINGS.md file. The resulting open-source project, BIOMA, is a lightweight Rust-core kernel with a Python interface designed to harden LLM prompts by stripping secrets, compressing context, and detecting prompt-flood attacks. In testing, the tool reduced a 47,890-token session to 2,022 tokens with roughly 1.6 microseconds of latency, and a 32,317-token flood attack was dehydrated to 13 tokens in 0.6 microseconds with zero secrets leaked. BIOMA is released under the FSL-1.1-MIT licence, which makes the source auditable and auto-converts to full MIT after a set period.

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)

Log in to join the discussion and vote.

Log in

Related stories

0
ProgrammingDEV Community ·

How the N+1 Problem Silently Slows Laravel GraphQL APIs and How to Fix It

The N+1 query problem is a common performance issue in GraphQL APIs built with Laravel, where a single query triggers one additional database call for each item in a list. For example, fetching 20 articles with their authors can generate 21 separate SQL queries instead of just two. In REST APIs, eager loading with explicit 'with()' calls makes the problem visible in code, but in GraphQL the risk is structural because relations are resolved automatically per request. The Lighthouse package for Laravel addresses this by default through its standard relation directives such as @belongsTo and @hasMany, which batch all required keys and execute a single grouped query. However, the problem can resurface with custom resolvers for computed fields like average ratings, where naive implementations still execute one database query per list item.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Developer Builds Open-Source Firewall for AI Agents After Near-Miss Database Deletion

A developer built an open-source security gateway called Bastion Gateway after a research AI agent nearly dropped a production database table, stopped only by a 403 error. The tool was designed to address two unguarded layers in the developer's Agent OSI model: identity and governance. Bastion Gateway works by enforcing per-agent allowlists, stripping sensitive data from outbound payloads, holding destructive actions for human approval, and maintaining a signed, immutable audit log. Every agent decision is recorded in an ed25519-signed JSON log following the Compliance-as-Code evidence spec, intended as verifiable proof for auditors and compliance teams. The self-hosted version is free and available on GitHub under the Apache-2.0 license, with a hosted version offering additional features currently on a waitlist.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Single Responsibility Principle: Why Each Class Should Do Only One Thing

The Single Responsibility Principle (SRP), part of the SOLID design guidelines, states that a class should have only one reason to change. Developers often bundle multiple functions — such as validation, password hashing, database storage, and email sending — into a single class for convenience, which makes the code fragile and hard to maintain. When a change in one area, like updating an email template, risks breaking unrelated logic such as data validation, debugging becomes time-consuming and error-prone. SRP addresses this by splitting responsibilities into dedicated classes, each handling a single concern, which improves testability, clarity, and flexibility. The practical result is that swapping out one component, such as an email provider, no longer risks introducing bugs in unrelated parts of the codebase.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

AWS Engineer Tests IAM and Lake Formation Auth Differences for Glue and S3 Tables Endpoints

AWS Community Builder Aki conducted hands-on tests to verify how IAM policies and Lake Formation grants affect access to the Glue REST Catalog and S3 Tables Iceberg REST endpoints. According to AWS documentation, the Glue endpoint requires both IAM and Lake Formation authorization, while the S3 Tables endpoint relies on IAM alone. The experiment used a matrix of four permission combinations — varying IAM and Lake Formation grants independently — to observe how each endpoint responded with HTTP 200 or 403 status codes. CloudTrail logs were also examined to confirm whether the lakeformation:GetDataAccess API call was triggered, revealing the internal authorization flow for each endpoint. The findings aimed to validate whether Lake Formation is genuinely enforced on the Glue side and confirm the simpler IAM-only path for S3 Tables.

0 comments Read more at DEV Community