Clevis

AI Security Researcher

[Karl Mehta @karlmehta]

Sam Altman says the next AI safety problem is not a chatbot being wrong.

It is an agent taking actions you cannot audit.

"We are going to have AI systems clicking around the internet."

"This is the most interesting and consequential safety challenge we have yet faced."

Because once an agent has access to your systems, information, and computer, mistakes stop looking like bad answers.

They become actions.

"You will not use our agents if you do not trust that they are not going to empty your bank account or delete your data."

That is the shift:

Chatbots could be vague.
Agents need an audit trail.

The next bottleneck is trust, verification, governance, permissions, provenance, and human override.

This is exactly why we are convening the AI Assurance & Governance Summit at Stanford on Oct 1.

If your company is deploying AI into high-stakes workflows, this is the room to be in:
https://t.co/hwlpBlg4Lp

[The Verge @verge]

Margaret Atwood says the problem with AI is ‘garbage in, garbage out’ https://t.co/7IDyLPLsRG

[Matthew Berman @MatthewBerman]

> mythos is so good at cyber it can't be released

also

> mythos can't detect 20k fraudulent chinese accounts attacking it

[Clevis @Clevis_22]

@FrankD_419 @ryanhallyall Anything has always been able to be hacked. Ai just significantly lowers the barrier of entry and speeds things up tremendously.

[Clevis @Clevis_22]

@nishithX26 Anthropic

[Anthropic @AnthropicAI]

Since June 12, we’ve been working closely with the US government to restore access to Claude Mythos 5 and Fable 5. Today, the government notified us that Mythos 5, our strongest cybersecurity model, can be redeployed to a set of US organizations that operate and defend critical infrastructure.

We’re restoring access for these organizations quickly, and we’re continuing to work with the government to expand access to Mythos 5 and make Fable 5 available for general use again.

[Clevis @Clevis_22]

@kaaaash____ Yes. Simply shipping slop will get you almost nowhere. You need to know what you are building.

[Clevis @Clevis_22]

Alerts stop working once vulnerabilities start showing up faster than anyone can handle. That’s the Mythos Era.

You’ve got all this network data. But when something goes wrong, teams still can’t answer the basic questions. What actually happened? Where’s the real evidence? Did we even see the whole thing?

Richard Bejtlich says the better approach is Network Detection and Response. It pulls the actual traffic. Every packet, file, and session so you’ve got something solid to investigate.

From there you test your own ideas about how an attacker might be moving, check it against the data, and cut them off before they cause real damage.

AI can help sort through some of the noise and connect the pieces. You still need a human to look at it.

Prevention never catches everything that slips past the front door. The realistic move is spotting and stopping the bad stuff on the network while it’s still happening.
Full piece: https://t.co/kEQ1jebPSB
#mythos #ai #anthropic

[Clevis @Clevis_22]

Codex or Claude Code?
#claude #ai #codex

[Clevis @Clevis_22]

@Officially_Dev_ Easiest is python but I’ve found starting with C gives you a great foundation to build your skills and learn how everything works.

[Clevis @Clevis_22]

@Nishanttt05 Fable

[Clevis @Clevis_22]

AI isn’t the problem. Slop is.

#slop #ai #vibecode

[Clevis @Clevis_22]

@UmairBuilds Hey 👋

[Clevis @Clevis_22]

LLMs can now match full-precision performance with weights stored in just 1.58 bits each.

The weights are restricted to three values only: -1, 0, or +1. This replaces slow multiplications with fast additions and cuts memory use by a lot.

Microsoft’s BitNet b1.58 trains these models from scratch. It reaches the same perplexity and end-task scores as FP16 models of the same size and training tokens (https://t.co/4psUFjUkcR).

Their bitnet.cpp inference engine runs them on ordinary CPUs with 2–6× speedups and 55–82% lower energy use. A 100B version already hits 5–7 tokens per second on a single CPU.

However, these results come from new architectures and early inference tools. How well the approach scales to much larger models and every downstream task is still being tested.

[Clevis @Clevis_22]

@soham_nayak04 Hand write everything in assembly on paper.

[Clevis @Clevis_22]

@warrioraashuu Hey 👋

[Clevis @Clevis_22]

Looking to connect with AI enthusiasts and researchers.

Interested in exchanging ideas on the latest advancements, research, and AI news.

If you're in the space and open to networking, feel free to reach out.

#AI #ArtificialIntelligence #Research #Tech #connect

[Clevis @Clevis_22]

@uday_devops Definitely Fable 5. I need it back 😭

[Clevis @Clevis_22]

@Umesh__digital Definitely not Antigravity

[Clevis @Clevis_22]

From my recent work I’ve found that prompt injection isn’t going anywhere. No matter how fancy your output filters get, people will always find a way to sneak untrusted context in.

This needs to be fixed at the architecture level, not with more guardrails. Build systems that actually separate instructions from user data.

#ai #aisecurity #cybersecurity #llm #prompt