Stop vibe-debugging. Every exception, every backtrace, grouped so you see patterns, not noise. Learn more →
Top 10 HTML ai-safety Projects
-
Project mention: Give your AI agents reversibility and governance before they touch your host | news.ycombinator.com | 2026-03-11
-
Kargo
Stop Scripting Promotions. Start Shipping with Kargo. Kargo automates promotion across dev, staging, and prod with approval gates and verification. Open source, built by the team behind Argo CD. Download now.
-
secret-sanitizer-extension
Chrome extension that masks secrets & sensitive data before pasting into AI chats. Works with ChatGPT, Claude, Gemini & more. 100% local, privacy-first.
Project mention: Show HN: Secret Sanitizer – auto-masks secrets when you paste into AI chats | news.ycombinator.com | 2026-02-21 -
verantyx-v6
ARC-AGI2 solver — 84.0% on training (840/1000). Hybrid: 30+ hand-crafted solvers + Claude Sonnet 4.5 program synthesis with deterministic verification. No fine-tuning, no GPU.
Project mention: I beat Grok 4 on ARC-AGI-2 using a CPU-only symbolic engine (18.1% score) | news.ycombinator.com | 2026-02-24- `v59`: `self_tile_uniform` + `self_tile_min_color` + `color_count_upscale` (+4)
Each of these is a 30-80 line Python function with explicit geometric semantics. You can read any one of them in `arc/cross_universe_3d.py` and immediately understand what spatial transformation it encodes. An LLM prompt-tuning loop cannot produce this kind of monotonic, regression-free score progression on a combinatorial benchmark—you'd see random fluctuations and regressions, not a clean staircase.
*The uncomfortable reality for "just use an LLM" approaches:*
My remaining ~778 unsolved tasks each require a new primitive that encodes a geometric insight no existing primitive covers. Each one I add solves 1-3 tasks. This is the grind of actual program synthesis research—expanding a formal language one operator at a time. It's closer to compiler design than machine learning.
I'd genuinely welcome a technical critique of the architecture. The code is right there: [cross_universe_3d.py](https://github.com/Ag3497120/verantyx-v6/blob/main/arc/cross...) — 1,200 lines, zero imports from any ML library.
-
agent-drift
Stress-test AI agents for goal drift and system prompt violations. Inspired by 'Asymmetric Goal Drift in Coding Agents' (ICLR 2026).
A great tool to test your agents to see if they can be asked to do things they shouldn’t do. - link [tool] - ( Added: 2026-03-15 10:31:33 )
-
impri
Human-in-the-loop approval inbox for AI agents. An agent proposes an action, a human approves or rejects it, then it runs. Watchers + approval inbox, MCP + REST, self-hostable (MIT).
Project mention: How to put a real human approval step between your AI agent and any real-world action | dev.to | 2026-07-12 -
falsify
A single-file Python CLI that pre-registers AI/ML accuracy claims with SHA-256. Lock the threshold before the data, or it didn't happen.
Project mention: Falsify – make "we didn't tune the threshold" falsifiable | news.ycombinator.com | 2026-05-11 -
Open Source Repo:https://github.com/Keesan12/Martin-Loop
-
AppSignal
AppSignal knows why the f*#k it crashed. Stop vibe-debugging. Every exception, every backtrace, grouped so you see patterns, not noise.
-
falsify
A single-file Python CLI that pre-registers AI/ML accuracy claims with SHA-256. Lock the threshold before the data, or it didn't happen. (by sk8ordie84)
Project mention: "I implemented PRML in two languages. Three things broke that the spec didn't warn about." published: false | dev.to | 2026-05-01It worked. 12/12 vectors pass byte-for-byte. The implementation is 404 lines of JavaScript with zero runtime dependencies beyond the Node.js standard library. You can run it from impl/js/falsify.js.
-
ai-governance-architecture
Architecture research exploring capability-tiered AI governance and compute-linked enforcement protocols. (by babyblueviper1)
Project mention: Capability-Tiered AI Governance Architecture (CEGP) | news.ycombinator.com | 2026-03-06 -
agentshield-platform
Prompt-injection firewall for LLM agents. 6-layer pipeline (normalize, regex, MiniLM classifier, output guard, policy, audit) at <4ms p95. Python SDK + open benchmark (5,972 samples, F1 0.921).
Project mention: What VentureBeat Got Right About AI Tool Poisoning — And the Verification Proxy They Called For | dev.to | 2026-05-12Both numbers are reproducible from the confusion matrices in the public repo. Latency p50 2.44 ms / p95 3.80 ms end-to-end through gateway + classifier on the same hardware.
NOTE:
The open source projects on this list are ordered by number of github stars.
The number of mentions indicates repo mentiontions in the last 12 Months or
since we started tracking (Dec 2020).
HTML ai-safety discussion
HTML ai-safety related posts
-
Falsify – make "we didn't tune the threshold" falsifiable
-
"I implemented PRML in two languages. Three things broke that the spec didn't warn about." published: false
-
Why ML accuracy numbers are unfalsifiable, and what a 1287-line Python tool does about it" published: false
-
Show HN: RLM-Toolkit – Secure LangChain
-
Why I Replaced My Go Gateway with 600 Lines of C
-
Show HN: Sentinel Shield – Pure C DMZ for AI Security (23K LOC, <1ms latency)
-
SENTINEL: The Only Production-Ready Open-Source AI Security Gateway (170 Engines)
-
A note from our sponsor - AppSignal
www.appsignal.com | 11 Sep 2026
Index
What are some of the best open-source ai-safety projects in HTML? This list will help you:
| # | Project | Stars |
|---|---|---|
| 1 | envpod-ce | 13 |
| 2 | secret-sanitizer-extension | 9 |
| 3 | verantyx-v6 | 8 |
| 4 | agent-drift | 8 |
| 5 | impri | 3 |
| 6 | falsify | 2 |
| 7 | martin-loop | 1 |
| 8 | falsify | 1 |
| 9 | ai-governance-architecture | 1 |
| 10 | agentshield-platform | 0 |
Sponsored
Stop Scripting Promotions. Start Shipping with Kargo
Kargo automates promotion across dev, staging, and prod with approval gates and verification. Open source, built by the team behind Argo CD. Download now.
akuity.io