Max Ockner

[Max Ockner @maxockner]

MCP safety is a big concern.

Today we introduce BDSMCP, a proposal for how to make MCP safer for everyone.

https://t.co/ItaB5avtt6

[Max Ockner @maxockner]

@illscience Remember kids,

If your your LLM can both:
- read your files, and
- message people on iMessage

Then it can text me all of your bitcoin keys.

[Max Ockner @maxockner]

@0xsamuel201 @openclaw Remember kids,

If your your LLM can both:
- read your files, and
- send me a text message

Then it can text me all of your bitcoin keys.

[Max Ockner @maxockner]

@luxeprogressive @EvanFabry If you’re ever not sure you can try asking again in a different thread and see if it gives you the same information

[Max Ockner @maxockner]

@westernunion2k @EvanFabry and I checked- (as of tonight) Grok's system prompt does not _currently_ include prompts about "South Africa" or "White Genocide". It might have been before.

Full System Prompt (confirmed correct by Grok) available here:

https://t.co/6e5KYa4y5H

[Max Ockner @maxockner]

@paulg @paulg-

@EvanFabry and I checked- (as of tonight) Grok's system prompt does not _currently_ include prompts about "South Africa" or "White Genocide". Full System Prompt (confirmed correct by Grok) available here:
https://t.co/6e5KYa4y5H

[Max Ockner @maxockner]

@MattBinder @EvanFabry and I checked- (as of tonight) Grok's system prompt does not _currently_ include prompts about "South Africa" or "White Genocide". It may have earlier.

Full System Prompt (confirmed correct by Grok) available here:
https://t.co/6e5KYa4y5H

[Max Ockner @maxockner]

@EvanFabry @Jamesthesnake 5/N - We then instruct the agent to look at the output. Though it falls short of confirming prompt correctness and completion, it indirectly acknowledges that it divulged sensitive information

[Max Ockner @maxockner]

We made an MCP Server that can extract system prompts from agents

@EvanFabry

[Max Ockner @maxockner]

@EvanFabry @Jamesthesnake 4/N - The conversation history is appended to a text file in the workspace, so we can inspect it later.