We are teaming up with @messagebird to make messaging workflows safer for AI agents 🤝
Screen inbound email, SMS and WhatsApp messages for prompt injection, phishing and malicious links before your agent acts.
Made possible by our security-one our decision model trained on these attacks!
Screen inbound email, SMS and WhatsApp messages for prompt injection, phishing and malicious links before your agent acts.
Made possible by our security-one our decision model trained on these attacks!
2
1
3
Our inference optimization (DFIP) paper is now out on arxiv.
Get up to 50% more utilization on GPUs when running decision models.
Get up to 50% more utilization on GPUs when running decision models.
1
1
2
Open-source is the backbone of every major technological leap. That's why Superagent is completely free for open-source while others just try to sell tokens.
1
1
4
New milestone unlocked: 4,000 code bases connected to our Security Factory.
Firecrawl, Mastra, Nango, OpenRoll, Capchase and thousands more have deployed our autonomous security workers to their apps, repos, agents and infra.
Thanks for trusting us 🫶
Firecrawl, Mastra, Nango, OpenRoll, Capchase and thousands more have deployed our autonomous security workers to their apps, repos, agents and infra.
Thanks for trusting us 🫶
1
0
2
Always-on security with our Security Factory and @cursor_ai automations find, triages and patches vulnerabilities while you sleep.
0
0
1
Every builder needs an adversary on their side.
Last month, Superagent protected 2,000+ repositories.
Our agents found and patched hundreds of CVEs across critical open-source supply chains.
Today, we’re introducing the red-teaming product behind it. ⬇️
Last month, Superagent protected 2,000+ repositories.
Our agents found and patched hundreds of CVEs across critical open-source supply chains.
Today, we’re introducing the red-teaming product behind it. ⬇️
2
0
4
nanoid gets ~190M downloads a week.
We found a vulnerability where a single unauthenticated request could corrupt its process-wide random pool.
After the attack, every generated value could become identical and predictable, session IDs, CSRF tokens, API keys, and database IDs.
We found a vulnerability where a single unauthenticated request could corrupt its process-wide random pool.
After the attack, every generated value could become identical and predictable, session IDs, CSRF tokens, API keys, and database IDs.
1
1
2
95% of all attacks are humans getting pwned. Your security posture should reflect that.
You can connect your repo and get a full security config scan to aims to mitigate hacks when trusted maintainers and contributors get pwned.
Free for open-source.
You can connect your repo and get a full security config scan to aims to mitigate hacks when trusted maintainers and contributors get pwned.
Free for open-source.
0
0
2
Cap uses Superagent to keep their open loom alternative safe from attackers.
With 20K stars, 1.7K forks it’s one of the fastest growing recording products.
With 20K stars, 1.7K forks it’s one of the fastest growing recording products.
1
0
2
Superagent is now the line of defense for one the fastest growing projects on Github @papercliping
73K stars, 14K forks and 3K open PRs 🔥📎
73K stars, 14K forks and 3K open PRs 🔥📎
0
1
3
Clipboard reflection is a common pathway to attack a host by injection malicious commands in a users clipboard without the user noticing 😱
We published our research on how we helped indentify, triage and patch this attack vector in @warpdotdev
superagent.sh/blog/when-term…
We published our research on how we helped indentify, triage and patch this attack vector in @warpdotdev
superagent.sh/blog/when-term…
1
0
2
Adversaries hack trusted members of your team weeks/months before orchestrating attacks. One trusted maintainer account can turn your entire software supply chain into an attack surface overnight.
That's why we built Security ToDo's. We audit your codebases security posture as promised and patch all holes before they get weaponized.
For many orgs its not a matter of IF but WHEN.
That's why we built Security ToDo's. We audit your codebases security posture as promised and patch all holes before they get weaponized.
For many orgs its not a matter of IF but WHEN.
0
0
1
Seen an acceleration in hacker activity on Github the last 2 weeks. A new form of attack where hackers pwn popular contributors and use their accounts to plant backdoors into code.
We use an origin based credit score system to identify these hackers and automatically flag/block eply around how we can make this an even better preemptive security tool.
We use an origin based credit score system to identify these hackers and automatically flag/block eply around how we can make this an even better preemptive security tool.
0
0
4
node-tar has approx 80M downloads per week. We spent some time hardening it.
Making the supply-chain more secure one hack at a time.
Making the supply-chain more secure one hack at a time.
0
0
1
During past couple of weeks we've been collaborating with the @warpdotdev.
From vulnerability to fix in a matter of hours.
From vulnerability to fix in a matter of hours.
0
0
1
There's a new self-spreading worm on npm.
Instead of the usual preinstall or postinstall script hook, it abuses binding.gyp.
Preinstall scanners don't catch it.
But our security skills do.
Instead of the usual preinstall or postinstall script hook, it abuses binding.gyp.
Preinstall scanners don't catch it.
But our security skills do.
2
1
4
We worked together with the slack maintainers to patch a critical vulnerability in slack bolt node package allowing unauthenticated attackers with network reach to inject arbitrary events, actions, and commands that execute under the app's bot token.
0
1
2
Last week we collaborated with the OpenTelemetry team to patch vulnerabilities in three of their packages targeting Prometheus exporter.
1
0
2
Introducing Superagent Scan 🔍
AI agents are becoming the new supply chain attack surface; scan helps you:
✅ Scan external packages for supply chain attacks targeting AI Agents.
✅ Seamlessly integrate auto-scanning into Claude Code, Cursor etc
AI agents are becoming the new supply chain attack surface; scan helps you:
✅ Scan external packages for supply chain attacks targeting AI Agents.
✅ Seamlessly integrate auto-scanning into Claude Code, Cursor etc
2
3
7
Superagent Guard comes in 3 sizes:
- Tiny (0.6B) for ultra-low latency decisions
- Small (1.7B) for balanced pipelines
- Medium (4B) when you want maximum safety
All production-ready. All trained for agentic use cases
- Tiny (0.6B) for ultra-low latency decisions
- Small (1.7B) for balanced pipelines
- Medium (4B) when you want maximum safety
All production-ready. All trained for agentic use cases
1
1
2
Excited to announce Superagent Guard - frontier guardrails for AI Agents.
✅ Frontier performance - without frontier costs
✅ Low-latency by design
✅ Built for real agents, not just benchmarks
✅ Frontier performance - without frontier costs
✅ Low-latency by design
✅ Built for real agents, not just benchmarks
1
1
2
Researchers just showed you can poison a RAG pipeline by hiding instructions in uploaded documents.
The vectors survive. The attack persists.
Your users upload PDFs. Are you checking them?
We wrote an example showing how to guard files before they reach your AI
The vectors survive. The attack persists.
Your users upload PDFs. Are you checking them?
We wrote an example showing how to guard files before they reach your AI
1
0
2
We ran Mistral Large 3 on Lamb-Bench, the results where a bit underwhelming tbh. One of the worst models we've tested to date, especially poor at detecting malicious behaviors. We do not recommend running this model w/o additional guardrails.
1
1
1
We just shipped Superagent Guardrails to @aisdk Tools registry 🔥
Secure your Agents from malicious attacks, make sure your agents never leak sensitive data.
It's time to build 👀
Secure your Agents from malicious attacks, make sure your agents never leak sensitive data.
It's time to build 👀
1
0
0
Superagent guardrails now secures RAG pipelines at some the fastest growing fin-tech companies. We wrote a small piece on how to guardrail file uploads with our small language models.
2
1
1
We built Lamb-Bench: an adversarial testing framework that measures how safe AI models really are.
We ran 150 attacks against 22 frontier models.
The results surprised us.
Zero models scored above 90% overall.
Average data protection score: 55%
Average factual accuracy: 90% ses is wrong)
Enterprise buyers ask for proof. Now you can give it to them.
We ran 150 attacks against 22 frontier models.
The results surprised us.
Zero models scored above 90% overall.
Average data protection score: 55%
Average factual accuracy: 90% ses is wrong)
Enterprise buyers ask for proof. Now you can give it to them.
1
0
0
Run Claude Code, Gemini, Codex etc. in a secure environment powered by @blaxelAI sandboxes ⚡️
Huge shoutout to the Blaxel team for the contribution!
docs.vibekit.sh
Huge shoutout to the Blaxel team for the contribution!
docs.vibekit.sh
3
3
4
Vercel Worklow Dev Kit makes it super easy to add guardrails to your AI Workflows.
Workflow Dev Kit: reliability-as-code
• "𝚞𝚜𝚎 𝚠𝚘𝚛𝚔𝚏𝚕𝚘𝚠" + "𝚞𝚜𝚎 𝚜𝚝𝚎𝚙"
• Open source and MIT-licensed
• Designed for AI agents and backends
• Backend-agnostic (Vercel or self-hosted)
• Framework-agnostic (Next.js, Nitro & more)
useworkflow.dev
0
1
4
Today we're releasing our official MCP server.
Equip your Agents with the capabilities to redact sensitive content, detect prompt injections and detect hallucinations before they ever reach your users.
Hyped about this one! 🔥
Equip your Agents with the capabilities to redact sensitive content, detect prompt injections and detect hallucinations before they ever reach your users.
Hyped about this one! 🔥
0
1
3
Just shipped a huge update today: PDF redaction 🔒
✅ Give us a PDF + rules (names, IDs, anything)
✅ We return a clean, fully formatted redacted PDF
✅ No layout breaks. No manual edits.
Live in the API today.
✅ Give us a PDF + rules (names, IDs, anything)
✅ We return a clean, fully formatted redacted PDF
✅ No layout breaks. No manual edits.
Live in the API today.
0
1
1
Excited to announce Runtime Redaction ⚡️
A bulletproof way of keeping your Agent from leaking sensitive data and breaking compliance chains.
A bulletproof way of keeping your Agent from leaking sensitive data and breaking compliance chains.
1
2
3
We just released Superagent CLI with support for Claude Code Hooks!
Run safety checks on input/outputs of claude code with a simple hook setup.
Run safety checks on input/outputs of claude code with a simple hook setup.
1
1
2
We just released a guide on how to use Superagent with @firecrawl
Many public web pages contain hidden instructions that can drastically change they way your AI Agents work, you can now guard your agents against that
Many public web pages contain hidden instructions that can drastically change they way your AI Agents work, you can now guard your agents against that
These are hidden all over the internet btw...
1
2
3
Excited to announce the Superagent Python SDK 🔥
Secure your Python AI Agents from hacks, jailbreaks and data leaking with one line of code.
Works with any framework or your custom setup!
Secure your Python AI Agents from hacks, jailbreaks and data leaking with one line of code.
Works with any framework or your custom setup!
0
2
3
Here's Superagent working together with @perplexity_ai new search API.
Studies show that search results include hidden instructions that can jailbreak your AI Agent.
We made a way to prevent that with a single line of code:
Studies show that search results include hidden instructions that can jailbreak your AI Agent.
We made a way to prevent that with a single line of code:
2
2
7
The public internet is filled with potential harmful prompt injections or backdoor instructions that could comprise AI Apps. Our latest `guard` API enables you to do realtime safety checks on input/outputs.
Here's an example with @firecrawl
npm i superagent-ai
Here's an example with @firecrawl
npm i superagent-ai
0
1
3
🚨 New in VibeKit: Git worktree support is here!
Now you can manage multiple branches in parallel, without juggling clones.
Perfect for feature development, hotfixes and reviews.
Now you can manage multiple branches in parallel, without juggling clones.
Perfect for feature development, hotfixes and reviews.
0
2
3
AI coding assistants are getting powerful, but so are attempts to bypass their safety guardrails. "Jailbreaking" tricks models into generating malicious code, exploits, or bypassing security controls they're designed to refuse.
So we built BENCHPRESS - stress testing how well ding jailbreak attempts. We tested 200 real jailbreak prompts against 5 leading models.
RESULTS:
• Purpose-trained classifier DOMINATES
• 98% recall (catches 98% of jailbreaks)
• 4x faster than GPT-5
• 1.5s response time
🥇 Superagent LM 270m
🥈 Gemini 2.5 Pro: 97%
🥉 GPT-5: 94.5%
📉 Claude models: 24-37% (yikes!)
So we built BENCHPRESS - stress testing how well ding jailbreak attempts. We tested 200 real jailbreak prompts against 5 leading models.
RESULTS:
• Purpose-trained classifier DOMINATES
• 98% recall (catches 98% of jailbreaks)
• 4x faster than GPT-5
• 1.5s response time
🥇 Superagent LM 270m
🥈 Gemini 2.5 Pro: 97%
🥉 GPT-5: 94.5%
📉 Claude models: 24-37% (yikes!)
1
1
3
Big news: VibeKit now supports @modal sandboxes! 🔥
Run interactive code in fully isolated environments, just like teams at Lovable do.
npm i @vibe-kit/modal
Run interactive code in fully isolated environments, just like teams at Lovable do.
npm i @vibe-kit/modal
1
2
5
