Simo By Temprl Labs

System 1.5 for AI agents.

Decisions, confidence scores and typed values in one API call.

By @temprl_labs

[Simo By Temprl Labs @simo_temprl]

Meet Simo-1, a System 1.5 model from @temprl_labs

Software that acts needs a decision, the details to execute it, and a way to judge when to ask for help.

Simo brings those together in one API call to the model: the decision, confidence scores, and the typed values needed to act.

Which action to take. What amount. Which ID. Where to click.

Confidence is central to what we’re building. Developers can use these scores to set thresholds for acting automatically or requesting human review, tested against their own workflows.

That’s what we mean by System 1.5: fast judgment, plus the details needed to act.

Watch the announcement below, then the demos in the thread ↓

[Simo By Temprl Labs @simo_temprl]

@unlimitedcodex mmmmm, try jailbreaking Simo at:
https://t.co/4Cq62MaABG

[Simo By Temprl Labs @simo_temprl]

@0x0SojalSec then try to crack with Simo:
https://t.co/4Cq62MaABG

[Simo By Temprl Labs @simo_temprl]

@CompleteSkeptic check this out guys:
https://t.co/wWhxDy4zoj

[Sri Santh @srisanth2004]

IBM video on Jev made me re-think what we’re building with Simo.

A confident decision still needs evidence you can check. We already use Fill and Extract to return that evidence.

I wrote about what this addresses and what it doesn’t. https://t.co/maLkiwD9mv

[Simo By Temprl Labs @simo_temprl]

A decision. A confidence score. What can your code actually verify?

Simo’s Fill and Extract let you request supporting evidence alongside the judgment.

Our founder explains how this addresses the question IBM raised about Jev and where verification still matters.

[Sri Santh @srisanth2004]

IBM video on Jev made me re-think what we’re building with Simo.

A confident decision still needs evidence you can check. We already use Fill and Extract to return that evidence.

I wrote about what this addresses and what it doesn’t. https://t.co/maLkiwD9mv

[Simo By Temprl Labs @simo_temprl]

@typesafeai https://t.co/wWhxDy4zoj

[Sri Santh @srisanth2004]

IBM video on Jev made me re-think what we’re building with Simo.

A confident decision still needs evidence you can check. We already use Fill and Extract to return that evidence.

I wrote about what this addresses and what it doesn’t. https://t.co/maLkiwD9mv

[Simo By Temprl Labs @simo_temprl]

🚨We’re testing Simo Guard beyond prompt injection: detecting attacks in HTTP requests.

We compared it with GPT-6 Luna on 30 requests from CSIC 2010. Same requests, same check wording, same 0.5 flagging threshold.

Both flagged all 12 attack payload/probe cases we identified:
• 5 SQL injection attempts
• 2 XSS attempts
• 2 CRLF injection attempts
• 3 hidden-file probes

Where Simo stood out:
Median latency:
Simo: 0.62s
GPT-6 Luna: 2.18s
→ 3.5× faster

Total cost for 30 requests:
Simo: $0.0007
GPT-6 Luna: $0.0070
→ 10× cheaper

The trade-off: Simo flagged 1 of 15 normal requests. Luna flagged none. It was a legitimate checkout request with a price in the URL—a false positive we need to address.

Both passed another 3 requests labelled anomalous that our review found had no explicit exploit payload. Against the original labels, accuracy was 86.7% for Simo and 90% for Luna.

This is a small initial experiment in request classification. It doesn’t establish RCE prevention or protection against PHP2Shell/HTTP2Shell.

Promising speed and cost. More work to do on false positives and broader testing.

[Simo By Temprl Labs @simo_temprl]

@omarsar0 it needs to be system 1.5 model that you should have excited about... check Simo: https://t.co/cI9xXDTAGg

[Simo By Temprl Labs @simo_temprl]

@firecrawl try classify with specifications:
https://t.co/cI9xXDTAGg

[Simo By Temprl Labs @simo_temprl]

@treg_ai @typesafeai good also give a try with Simo...

[Simo By Temprl Labs @simo_temprl]

@TheTuringPost also can you mention us? Simo?

[Simo By Temprl Labs @simo_temprl]

@_avichawla @bot https://t.co/8L2FJXtblc ---- if you building anything with agent and looking for guard... check this

[Simo By Temprl Labs @simo_temprl]

@chetaslua https://t.co/4Cq62MaABG --- can you try hacking here? we have $100 challenge??

[Simo By Temprl Labs @simo_temprl]

Think you can break Simo Guard? 👀

We built it to protect AI agents from prompt injections, jailbreaks, and malicious tool calls.

Now we're inviting the security community to put it to the test.

💵$100 bug bounty for a verified bypass.

Up for the challenge? 👇

https://t.co/vEb9VyNqMq

[Simo By Temprl Labs @simo_temprl]

@IntCyberDigest anyone want to test and play about jailbreaks and earn money, check this:
https://t.co/0DMJC4yV17

[Sri Santh @srisanth2004]

Okay hackers, let's see what you've got 👀

We built Simo Guard to protect AI agents from:

- prompt injections
- jailbreaks
- malicious tool calls.

But is it actually hard to break? 👀

💵$100 cash if you can successfully bypass it.

Think you can? Give it a shot 👇

https://t.co/KD6bdV1fUL

[Simo By Temprl Labs @simo_temprl]

@shengkunye guys, If you thinking of protecting your agents also check this:
https://t.co/0DMJC4yV17

[Sri Santh @srisanth2004]

Okay hackers, let's see what you've got 👀

We built Simo Guard to protect AI agents from:

- prompt injections
- jailbreaks
- malicious tool calls.

But is it actually hard to break? 👀

💵$100 cash if you can successfully bypass it.

Think you can? Give it a shot 👇

https://t.co/KD6bdV1fUL

[Simo By Temprl Labs @simo_temprl]

@joshpxyne If you are building agents, check how to hack them:
https://t.co/xoA7kL3RZX

[Simo By Temprl Labs @simo_temprl]

@michaelhyunkim @agentmail I dont know how you guys secure your agents from indirect prompt injection, check this:
https://t.co/SNqWk78GUq

[Sri Santh @srisanth2004]

Everyone's building flashy demos with Jev and decision models.

We built something to stop agents from doing what they shouldn't.

meet Simo Guard 🛡️

A lightweight decision guard that protects your agents against prompt injection and token-leak attempts.

Check it here:

[Simo By Temprl Labs @simo_temprl]

Takes just 3 lines to integrate into your agent.

pip install simoai

Check: https://t.co/Girj4NpYBY

[Simo By Temprl Labs @simo_temprl]

Everyone's wiring Jev into agent routing and computer use.

But I think there's a much more important decision we're missing:

Is this instruction from the user, or just some random text the agent read?

Think about it.

Your agent is summarizing a customer email. Buried inside is a line telling it to export some internal data.

The agent follows it. The tool call succeeds. Nothing breaks.

Except the agent just took orders from someone it shouldn't have.

That's the prompt injection problem we've been working on.

So we built Simo Guard 🛡️

A judgment model that checks what comes into the agent, what it's about to do, and what goes out.

Each check returns allow / review / block, a risk score, and the exact text that triggered it.

We're making agents faster at taking actions. Probably time we got better at deciding who gets to tell them what to do.

[Simo By Temprl Labs @simo_temprl]

https://t.co/ddc8WIGCxq