Free course teaches developers to build AI agents with code-enforced safety limits

·3 views

A developer has launched Ship a Safe Agent, a hands-on TypeScript course on building AI agents that cannot take harmful actions. The author argues that prompts alone cannot prevent agents from issuing unapproved refunds or calling restricted tools, so safety must be enforced in code. The course requires only TypeScript, Node, async, JSON and git knowledge, and runs fully offline using scripted, deterministic models with no API key. Module 1, which covers the agent loop, is free and takes about 75 minutes, while the full course is priced at $149 until November 8, then $199. The course spans eight modules, including output validation, action gating, handling untrusted input, secrets management, cost controls and observability.

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)

TryHackMe Billing Room: Exploiting MagnusBilling Command Injection and fail2ban Escalation

This TryHackMe walkthrough targets the Billing machine, built on MagnusBilling, an open-source VoIP billing platform that runs alongside the Asterisk PBX. The initial foothold comes from an unauthenticated command injection flaw in the IcePay payment gateway integration, where an unvalidated parameter is passed into a shell context. The write-up explains how chained shell operators allow an attacker to fetch and execute a remote script for a reverse shell. The privilege escalation route abuses fail2ban, a security daemon, by manipulating its client binary to alter how banning actions are executed. The room is presented as a useful example of domain-specific attack surfaces in telephony software.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How AI Voice Models Are Trained: A Technical Walkthrough for Developers

A DEV Community article explains the typical pipeline for training modern AI voice and text-to-speech models, from data collection to inference. It outlines common public datasets such as LJSpeech, VCTK, and LibriTTS, and notes that a custom voice clone needs at least 30 minutes of clean recordings, with two to three hours yielding better quality. The piece covers preprocessing steps including resampling, silence trimming, loudness normalization, and grapheme-to-phoneme conversion, with sample Python code. It also compares two-stage architectures (acoustic model plus vocoder) such as Tacotron-2, FastSpeech, and VITS, and closes with a minimal PyTorch training loop for a FastSpeech-2 model. The author suggests that managed services like ElevenLabs can save considerable engineering time.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Why FAQ, HowTo and LocalBusiness schema on small WordPress sites often backfires

Plugins like Rank Math and Yoast often add FAQPage, HowTo and LocalBusiness markup that users never see, which creates clutter rather than value. The article notes that Google no longer shows FAQ or HowTo rich results for typical sites, though cleaning up the markup still matters for other systems and for local visibility. Problems include FAQ JSON-LD repeated sitewide, HowTo markup applied to pages that are not real procedures, and LocalBusiness details that disagree with the visible footer. The recommended approach is to keep FAQ markup only where visible Q&A is the page's main content, use HowTo only for genuine step-by-step guides, and maintain one accurate LocalBusiness entry. Duplicate markup should be removed and canonical tags used to settle which URL owns a topic.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Teams Split AI Agents Into Specialized Roles Under an Orchestrator

Developers are increasingly replacing single all-purpose AI agents with several specialized sub-agents, such as a planner, coder, and reviewer, coordinated by a central orchestrator. The shift is motivated by the difficulty a lone agent has with long, multi-skill tasks, where crowded context and compounding early errors degrade results. Smaller focused agents are easier to evaluate individually, making it simpler to trace which step caused a faulty output. The article outlines three patterns: orchestrator-worker, peer review pairs, and fixed pipeline handoffs. It recommends this approach mainly for complex tasks needing distinct skills, while noting that coordination overhead can outweigh the benefits for simpler work.

0 comments Read more at DEV Community