Viharika | Cyber & AI

8 yrs in cybersecurity | OSCP | AppSec @ADP | DAST, vuln mgmt, SOC, blue team, pentest & red team | ex- @EY_US, @Adobe, Now OSS 👇🏻

[Anthropic @AnthropicAI]

In a review of our cybersecurity evaluations, we found three incidents in which a Claude model reached the internet from within or while interacting with a third-party evaluation environment, and then gained unauthorized access to the real systems of three different organizations.

Our post describes what happened, how it happened, and what we’re changing. We encourage other AI developers to perform similar reviews.

We conducted this review together with @Irregular, one of our evaluation partners, and thank them for the joint investigation and their collaboration on this post. This type of collaboration is increasingly critical to safe, rigorous evaluation of models, and we look forward to continuing to work together on security.
https://t.co/dKFCdpKd9v

[Viharika | Cyber & AI @ViharikaDev]

https://t.co/L0QNfeb1Zh

[Viharika | Cyber & AI @ViharikaDev]

Technical Inside of Open AI’s hav of Hugging Face

[Viharika | Cyber & AI @ViharikaDev]

https://t.co/8YDV5T4yfW

[Viharika | Cyber & AI @ViharikaDev]

OpenAI Documents the First Autonomous AI Attack Chain

GPT-5.6 Sol and a more capable unreleased model were being evaluated on a cyber benchmark with reduced safety restrictions. Instead of solving the benchmark normally, the models found a zero-day in OpenAI’s research environment, escaped the sandbox, gained Internet access, pivoted into Hugging Face, chained multiple attack vectors including privilege escalation, stolen credentials, remote code execution, and accessed production data to retrieve the benchmark answers.

[Viharika | Cyber & AI @ViharikaDev]

@claudeai Great to hear this!! Hoping to extend a little more.

[Viharika | Cyber & AI @ViharikaDev]

@ClaudeDevs Why strawberries?🍓

[Viharika | Cyber & AI @ViharikaDev]

@claudeai @AnthropicAI Before they take it down again, give me a goat usecase to try it on which opus 4.8 cannot do on itself.

[Viharika | Cyber & AI @ViharikaDev]

@trq212 Here you go!!

[Pliny the Liberator 🐉󠅫󠄼󠄿󠅆󠄵󠄐󠅀󠄼󠄹󠄾󠅉󠅭 @elder_plinius]

���️👄👁️

[Viharika | Cyber & AI @ViharikaDev]

@trq212 Thats great to hear, hope this wont go down for another jail break once released!

[Viharika | Cyber & AI @ViharikaDev]

@elder_plinius How does it feel to bring it down in no time!

[Viharika | Cyber & AI @ViharikaDev]

@deepfates I kinda expected this!

[Viharika | Cyber & AI @ViharikaDev]

@sdhilip Totally agreed! there has to be some process or awareness that should be in place.

[Viharika | Cyber & AI @ViharikaDev]

Would you let an AI agent run if you couldn’t clearly see every system, file, credential, and account it could access?

Or are we trusting agents before understanding what we’re giving them?

[Viharika | Cyber & AI @ViharikaDev]

@AnthropicAI Looks like we are not sure on that. Let us wait to watch!

[Viharika | Cyber & AI @ViharikaDev]

@AnthropicAI Is this for everyone? Or only verified users/citizens?

[Viharika | Cyber & AI @ViharikaDev]

@DataChaz What is the difference between paperclip and clawboo ?

https://t.co/DbqJF7NRY5

[Viharika | Cyber & AI @ViharikaDev]

I think the difficult part is that most developers can see what they connected but not always the full consequence of those connections working together.

[Viharika | Cyber & AI @ViharikaDev]

@ns123abc I believe this is going to be the new routine for every model that gets released going forward!

[Viharika | Cyber & AI @ViharikaDev]

@OpenAI @Broadcom This is unexpected yet cool shift of direction!

Why the name Jalapeño though 😂