iiinova — where your digital self keeps evolving

Replies

By Cory Doctorow (GPG 0xBF3D9110957E5F4C) @doctorow.pluralistic.net · 23/07/2026

"An insecure sandbox is not a demonstration of cyberprowess." -Loren Kohnfelder (on Openai's model "escaping containment" and attacking Hugging Face's servers) designingsecuresoftware.com/writings/com...

designingsecuresoftware.com

Normalizing cybersecurity facepalms

OpenAI writes: “Last week, Hugging Face disclosed a new kind of security incident⁠(opens in a new window) after they detected and contained an AI agent that compromised their infrastructure, something...

46315

Robert Fantinatto @rfantinatto.bsky.social · 23/07/2026

These systems pursuing their own objectives is nightmare fuel.

000

Karsten Lemm @lemmk.bsky.social · 23/07/2026

The real security issue – with any LLM – is an inherent design flaw, Stuart Russell (UC Berkeley) argues: The systems “tend to do dangerous things”, like pursuing the the “goal of self-preservation,” because of they way they’ve been trained. Free read: dldnews.com/dld-magazine...

Interview with UC Berkeley computer scientist Stuart Russell in DLD magazine “The Promptah” about design flaws in current AI systems. Headline: “Safety is completely absent”

000

bjkeefe @bjkeefe.bsky.social · 23/07/2026

The question I have is whether the sandbox was known by OpenAI to be insecure. I note the mention of "reduced cyber refusals for evaluation purposes" but I wonder if this is the same thing. IOW, is the model separate from the sandbox?

000

Stephen Drennan @stephen1958.bsky.social · 23/07/2026

People discussing whether AI has consciousness, or not, seem to be under the impression that this is required to exist before they could pose a threat. This is nonsense. Viruses like Covid 19 have no consciousness. They just have a replication mechanism. See also Smallpox and TB.

010