MATT Security

The ExploitOps Platform. Continuously secure your products as AI and attack techniques evolve. Find new attack surfaces. Verify fixes.

[MATT Security @mattsecures]

Episode II: UNCENSORED.

A safety guardrail inside a model anyone can download lasted about a week.

Our film on what happened next.
#Cybersecurity #AI #Claude #GLM

[MATT Security @mattsecures]

8/ Both reports, every figure, and the film on the same story:
https://t.co/RE2WKZWmAW

[MATT Security @mattsecures]

1/ A working attack on a known Chrome flaw now costs $20.40.

@AnthropicAI's number, from its test of GLM-5.3, a free AI model from https://t.co/aaTUcqd0Ea that anyone can download. Eight hours of the model, twenty minutes of a person.

Here is what two reports measured.

[MATT Security @mattsecures]

7/ What to do:
→ shorten the time from a published flaw to a deployed patch
→ write down the rules your product must hold; check them on every release
→ treat a model's safeguards as one layer, because an open model's go wherever its weights go

[MATT Security @mattsecures]

@Nuke_cryptovn @OpenAI @claudeai @huggingface @Nuke_cryptovn here you go https://t.co/8UzkwgSexU 🙌

[MATT Security @mattsecures]

Episode II: UNCENSORED.

A safety guardrail inside a model anyone can download lasted about a week.

Our film on what happened next.
#Cybersecurity #AI #Claude #GLM

[MATT Security @mattsecures]

A short movie on the biggest AI security incidents of 2026, ft @OpenAI, @claudeai and @huggingface.

Let us know if we should do Episode II?

#AI #securitynews

[Subho Halder 👨🏽‍💻 @sunnyrockzzs]

1/ In July, @OpenAI's test agents found 14 Hugging Face logins left out in public.

For four days they ran code on @huggingface's servers. No human directed a single step.

We made a short film about it, "WARNING SHOT".

[MATT Security @mattsecures]

For those who haven't checked out Episode I, Warning shot, here's the film https://t.co/vOdjSSMvdU

[MATT Security @mattsecures]

Episode II: UNCENSORED.

A safety guardrail inside a model anyone can download lasted about a week.

Our film on what happened next.
#Cybersecurity #AI #Claude #GLM

[MATT Security @mattsecures]

10/ The Accord, explained in plain English, with the full page: https://t.co/NsYGZpc4Nt

[MATT Security @mattsecures]

1/ Google, Anthropic, Meta, OpenAI, xAI and Nvidia signed one page at the White House.

312 words. Four promises.

Promise three: an outsider must check each company's safety controls.

It does not say who. Here is what the White House Accord on Super Intelligence leaves out.

[MATT Security @mattsecures]

9/ So ask the four layers of yourself:

1. What must always be true?
2. Who inside checks it?
3. Who outside has proven it?
4. What does the board see?