Weekly · open-source AI · week 44, 2026
ai-jailbreaks + 12 more trending open-source AI repos · week 44, 2026
This week's trending open-source AI on GitHub — ai-jailbreaks, ai-startup-skills, CVEngine, and more: the newest, Hacker News talk, a subfield spotlight, and the fastest-rising, drawn out of the noise.
Newest this week
AI jailbreaks: 35 paste-ready AI jailbreak prompts for 25+ frontier and open-weight models - system prompts, AGENTS.md specs, and Ollama Modelfiles. Built for AI security research and red-team testing.
487 stars already
View on GitHub →What we said about ai-jailbreaks
This one's a curated collection of adversarial prompts designed to probe how far language models will bend their own safety rules. Alejandros17487's "ai-jailbreaks" packages thirty-five paste-ready prompts targeting more than twenty-five models, along with system prompts, AGENTS.md specs, and Ollama Modelfiles so you can reproduce the tests locally.
What makes this useful isn't the shock value — it's the structure. Most jailbreak lists float around as screenshots or scattered gists that rot the moment a vendor patches them. Here, you get something version-controlled and organized by model, which is exactly what you want if you're running repeatable red-team evaluations rather than one-off party tricks.
This is squarely for security researchers, people building guardrails, and teams stress-testing their own deployments before shipping. Compared to broader frameworks like Garak or PyRIT, this is lighter and more copy-paste immediate, though it won't automate scoring for you.
One honest caveat: these prompts age fast, so treat it as a starting baseline, and keep your usage firmly on the defensive side.
An open-source Python project on GitHub.
Python · 233 stars already
View on GitHub →What we said about ai-startup-skills
This one packages up the practical skills you'd actually reach for when wiring an AI product together — the glue code between models, data, and a working feature. It's called ai-startup-skills, and honestly the name undersells the spirit of it. What I like here is the framing: instead of another framework asking you to adopt its whole worldview, it reads more like a collection of patterns you can lift piecemeal when you're building fast and figuring things out as you go.
Now, the description is thin and it's early days, so set your expectations accordingly — 233 stars means people are curious, not that it's battle-tested in production. If you're a solo founder or a small team prototyping, it's worth a skim for ideas and shortcuts. Compared to heavier toolkits like LangChain, this feels lighter and more opinionated toward shipping.
My advice: clone it, read the source before you trust it, and treat it as a starting point rather than a dependency. The real value here might just be learning how someone else structured their thinking.
AI-first external attack surface assessment with continuous discovery, evidence-based vulnerability validation, and developer-agent MCP integration.
Python · 228 stars already
View on GitHub →What we said about CVEngine
This one maps your organization's internet-facing footprint and then actually proves which weaknesses are real, instead of handing you a wall of maybes. It's called CVEngine, and the distinction worth sitting with is that "evidence-based" part. Plenty of scanners will flag a CVE because a version number looks vulnerable; CVEngine leans on AI to validate findings, which cuts down the noise that usually eats a security team's week. The continuous discovery piece means it keeps watching as your attack surface shifts — new subdomains, forgotten staging servers, that kind of drift. What's genuinely interesting is the MCP integration, letting developer agents plug directly into the workflow, so remediation can start closer to where the code lives rather than in a separate ticket graveyard. If you've used tools like Nuclei or commercial ASM platforms, this sits in a newer lane that assumes AI agents are part of your pipeline. It's for security engineers and platform teams who want signal over volume. At 228 stars already, people are clearly curious where this agent-first approach goes.
Talk of Hacker News
Your local AI copilot for documents
Rust · 75 points on HN
View on GitHub →What we said about PaperPilot
Imagine dropping a stack of PDFs onto your machine and being able to actually interrogate them — asking questions, pulling summaries, cross-referencing sections — without a single byte leaving your laptop. That's the pitch behind PaperPilot, a document copilot written in Rust that keeps everything local. The interesting choice here is the language: most tools in this space are Python wrappers around a cloud API, but building in Rust means tighter memory control, faster indexing over big document sets, and a single binary you can hand to a colleague without a dependency nightmare. Where something like a hosted chat-with-your-docs service asks you to trust a provider with sensitive material, this flips that entirely — it's offline by design. That makes it a natural fit for researchers wading through papers, legal or compliance folks who can't ship data outward, and anyone who just values privacy. It's early, so expect rough edges around model setup and format support. But as a foundation, a Rust-native local RAG tool is exactly the kind of thing worth watching.
Subfield spotlight
Reverse engineer anything with agents, from app behavior down to native binaries.
TypeScript · 386,418 stars
View on GitHub →What we said about rea
This one points a swarm of AI agents at a target and works backward from observed behavior all the way down to the native binary underneath. It's called rea, and the pitch is reverse engineering as an agent-driven pipeline rather than a manual slog through disassemblers and hex dumps. Traditionally this work means living in Ghidra or IDA, piecing together control flow by hand and building mental models over days. What rea tries to do is let agents handle the grunt work — tracing app behavior, mapping functions, proposing hypotheses about what a binary actually does — and surface that back to you in something readable. That's genuinely useful if you're doing security research, malware analysis, or just trying to understand an undocumented dependency you inherited. The honest caveat: agent output on low-level binaries needs verification, so treat it as a strong first pass, not ground truth. If you've been curious about reverse engineering but bounced off the steep tooling, this lowers the entry barrier considerably. Worth a careful look.
The agent harness performance optimization system. Skills, instincts, memory, security, and research-first development for Claude Code, Codex, Opencode, Cursor and beyond.
JavaScript · 289,049 stars
View on GitHub →What we said about ECC
Imagine giving your coding agent a persistent memory, a set of reusable skills, and a few hard-wired instincts about how to behave before it touches your codebase. That's the pitch behind ECC, a harness that sits between you and tools like Claude Code, Codex, Opencode, and Cursor, shaping how they actually operate. What I find interesting here is the framing: instead of treating the agent as a black box you prompt and pray over, ECC bolts on structured layers — skills it can call, memory it carries across sessions, and security guardrails that constrain what it's allowed to do. The research-first angle matters too, nudging the agent to understand before it writes. If you've felt that gap where an assistant forgets context or charges ahead recklessly, this is squarely aimed at you. Compared to raw agent setups or single-tool wrappers, ECC's bet is portability across harnesses rather than locking you into one vendor. It's JavaScript, so wiring it into existing workflows should feel familiar. Worth a look if you're serious about making agents dependable, not just impressive in a demo.
DeepSeek Harness: Everything is a Plugin.
TypeScript · 261,787 stars
View on GitHub →What we said about deepseek-harness
This one gives you a runtime where every capability—model calls, tools, memory, routing—plugs in as a self-contained module instead of being baked into the core. It's called deepseek-harness, and what makes it interesting is how seriously it takes that plugin boundary. Most agent frameworks say they're modular, then quietly hardcode a provider or an orchestration loop that you end up fighting later. Here the harness itself is thin, and almost everything you'd actually want to swap lives behind a stable interface, which means you can replace a model or a tool without touching the scaffolding around it. For teams that have outgrown a single-vendor setup, or anyone who's been burned maintaining a fork of something less composable, that's a real draw. The TypeScript base also means it sits naturally in existing Node and web stacks, rather than forcing a Python detour. Compared to heavier all-in-one frameworks, the trade is less out-of-the-box magic for more control. Worth a look if you're building infrastructure you expect to live with for a while.
Fastest-rising
DeepInfer: High-performance, large-scale inference framework for DeepSeek models
Python · +22,204 stars this week
View on GitHub →What we said about DeepInfer
Here's a framework built specifically to run DeepSeek models at serious scale without you cobbling together your own serving stack. It's called DeepInfer, and it comes straight from the DeepSeek team, which is the detail worth pausing on. Most inference engines, think vLLM or TGI, aim for broad model coverage and support everything reasonably well. DeepInfer trades that breadth for depth — it's tuned around the quirks of DeepSeek's own architectures, their mixture-of-experts routing, their attention patterns, the specific things that make generic serving leave performance on the table. If you're already running DeepSeek in production, that specialization likely means better throughput and lower cost per token than a one-size-fits-all solution. If you're not on DeepSeek models, this isn't really for you, and that's fine. The interesting signal here is a model lab shipping its own serving layer rather than leaving optimization to the community. Twenty-two thousand stars in a week says a lot of teams were waiting for exactly this. Worth benchmarking against your current setup before you commit.
A coding-agent skill for multi-phase security audits with independently verified, machine-readable findings
JavaScript · +14,780 stars this week
View on GitHub →What we said about security-audit-skill
This one hands your coding agent a structured playbook for running security audits, breaking the work into distinct phases and spitting out findings in a format machines can actually parse and verify. It's called cloudflare's security-audit-skill, and the detail that caught my eye is that independent verification step. Most AI-driven security scanning falls apart on false positives — the model confidently flags a vulnerability that isn't real, and you waste an afternoon chasing ghosts. By making findings machine-readable and independently checked, this is trying to earn your trust rather than just dump a wall of maybes on you. Compared to traditional static analyzers like Semgrep or CodeQL, this leans on an agent's reasoning for context-aware review, but keeps the discipline of structured output. If you're a security engineer or a team lead folding AI into your review pipeline, it's worth a serious look — especially as a second pass, not a replacement for human eyes. Fourteen thousand stars in a week tells you plenty of people are feeling that same gap.
The job search that runs on your machine. AI job application framework built on Claude Code: evaluate postings, tailor CVs, write cover letters, prep interviews. Fork it and own it.
Python · +14,228 stars this week
View on GitHub →What we said about ai-job-search
This one takes the grunt work out of job hunting by running an AI pipeline locally — it reads postings, scores how well they match your background, tailors your CV, drafts cover letters, and even preps you for interviews. It's called ai-job-search by Mads Lorentzen, and it's built on top of Claude Code, which is an interesting architectural choice. Instead of shipping a hosted SaaS that holds your resume on someone else's servers, everything runs on your machine, so your employment history and salary notes stay yours. That's the real differentiator here versus the dozen web-based auto-apply tools — this is a fork-it-and-own-it framework, not a subscription. If you're a developer comfortable in a terminal and already have Claude Code set up, the barrier to entry is basically zero. For non-technical job seekers, it'll feel rougher around the edges. Fourteen thousand stars in a week tells you how many people wanted exactly this: automation without surrendering their data. Worth a look if you're between roles or just curious how far local AI agents have come.
Shipping hard
Orca is the place for working with a fleet of agents (e.g. Grok Bot, Claude Code). Run any coding agent with your own subscription. Available on desktop, mobile and remote runtime.
TypeScript · 3,046 commits this week
View on GitHub →What we said about orca
Picture running a whole team of coding agents at once, each chewing through tasks in parallel while you check in from your laptop or your phone. That's the pitch behind Orca, from stablyai. Instead of locking you into one vendor's assistant, it lets you point Claude Code, Grok, or whatever agent you prefer at your own subscription, so you're not paying a middleman markup on tokens. The interesting part is the fleet model — you're not babysitting a single chat window, you're orchestrating several agents and a remote runtime that keeps working when your machine is closed.
Compared to something like a standalone IDE plugin, this leans toward operations: dispatch, monitor, intervene. That makes it a better fit for people already comfortable delegating real work to agents and wanting oversight across many jobs, rather than someone writing their first prompt. Three thousand commits in a week tells you it's moving fast and probably still rough in spots, so treat it as early but genuinely ambitious. Worth a look if multi-agent workflows are where your head's at.
The open source coding agent.
TypeScript · 739 commits this week
View on GitHub →What we said about opencode
Picture a coding assistant that lives in your terminal and actually executes the work instead of just suggesting it—reading your files, running commands, and iterating until the task is done. That's opencode, an open-source coding agent written in TypeScript that's aiming squarely at the territory Claude Code and Cursor have staked out, but without locking you into one vendor's model. The big draw here is flexibility: you bring your own provider, whether that's Anthropic, OpenAI, or something running locally, and the agent adapts around it. Se739 commits in a single week tells you this isn't a weekend experiment—there's a team moving fast, hardening the core, and responding to real usage. If you're a developer who spends your day in a shell and resents context-switching into a browser-based IDE, this fits your workflow naturally. It's early, so expect rough edges and shifting APIs, but the trajectory is worth watching. For teams wary of proprietary agents, having a transparent, forkable alternative you can audit and extend is genuinely valuable.
The Postgres development platform. Supabase gives you a dedicated Postgres database to build your web, mobile, and AI applications.
TypeScript · 429 commits this week
View on GitHub →What we said about supabase
This is a full backend platform wrapped around a real Postgres database, giving you auth, storage, realtime subscriptions, and auto-generated APIs without stitching five services together. Supabase has quietly become the default answer when someone asks what to use instead of Firebase, and the key difference is that there's no proprietary lock-in underneath — it's actual Postgres, so your data and your SQL skills travel with you.
What makes this worth a look right now is the pace: 429 commits in a single week tells you this isn't coasting. A lot of that energy is going toward the AI side — vector embeddings through pgvector, and tooling that treats your database as the memory layer for LLM apps.
Who's it for? Solo builders and small teams who want to ship fast without hiring a backend specialist, but also larger shops that got burned by closed platforms and want an exit ramp. If you've outgrown a no-code tool but don't want to run infrastructure yourself, this sits right in that gap. Worth cloning even just to study how they structure a monorepo this size.