SShortSingh.
TopTechnologyAI & MLStartupsProgrammingScienceBusinessFinanceCrypto & Web3World
Back to feed
Programming

New Tool Tests AI Security Against Dormant Prompt Injection Attacks

0
·1 views

A developer has released deny-probe v0.4.0 to test AI security systems against trigger-based prompt injections. These attacks remain dormant in context until specific triggers activate them, making traditional denylist defenses less effective. Research shows trigger-based injections succeed 43-83% of the time compared to 3% for direct instruction attacks. The tool tests whether security rules can block actions once dormant payloads are activated. It runs as a Python package with deterministic testing suitable for continuous integration systems.

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 ·

Hyper3D and SupaVoxel AI model services tested, show varied cost and output results.

An independent test compared AI 3D model generation services Hyper3D and SupaVoxel. Using the same source image, the tester attempted to generate a 3D model of a cat. Hyper3D's Gen-2.5 preview consumed 0.5 credits but then required a subscription to download, resulting in no usable file. The tester successfully generated a textured 2K model using Hyper3D's Gen-1.5 for 0.5 credits, while SupaVoxel delivered a 4K textured model for 3 credits. The test concluded that each service offered different value propositions based on output quality and credit cost.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Developer creates offline AI gardening planner for Hacktoberfest 2026 challenge

A developer has built GardenPlanner AI, an offline-first gardening tool for Hacktoberfest 2026's Open Source AI Challenge. The command-line tool generates planting plans using local climate data, seasonal adjustments for hemispheres, and companion planting guidelines without requiring internet connectivity. It operates with a built-in plant database and offers optional local AI integration via Ollama for natural language queries. The project is designed for community contributions through Hacktoberfest, encouraging additions to its climate tables and plant databases.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Subagent-tax 0.3 update corrects billing model, revealing significantly higher token costs

Developer hahahahahahahahah6 released version 0.3 of the subagent-tax auditing tool, which corrects a flawed billing calculation. The previous version undercounted token usage by only charging for a fixed preamble once per task, rather than on every API call and subagent spawn. Real-world data showed the old model underestimated costs by a factor of three, with subagents consuming nearly half of all tokens in one analysis. The update introduces a detailed billing ledger and shares its framework with the related mcp-tax auditing tool. The Python package is available on PyPI under an MIT license.

0 comments Read more at DEV Community