Friday Squid Blogging: The Colossal Squid
Long article on the colossal squid.
Long article on the colossal squid.
lurker • February 10, 2025 12:31 PM
@Slive Robinson
It’s easy: first you sack the people doing the vetting …
Clive Robinson • February 12, 2025 8:15 PM
@ Bruce,
“All is not well on the Safe AI Front”
https://arstechnica.com/ai/2025/02/us-and-uk-refuse-to-sign-ai-safety-declaration-at-summit/
Clive Robinson • February 13, 2025 6:37 AM
@ Bruce, ALL
Another AI and crypto project.
Apparently, “Textcoder is a proof-of-concept tool for steganographically encoding secret messages such that they appear as ordinary, unrelated text.”
(Which is something I’ve talked about off and on on this blog for a number of years, including the past few days using “Perfect Secrecy” encryption and a Code Book of phrases. So people can look back if they want the “basic and secure” idea).
This time however the system rather than use a traditional “code book” of phrases or words, the project,
https://github.com/shawnz/textcoder
Uses a current AI LLM system to get “Statistical weighting”.
To quote the project page,
“It works by taking the secret message, encrypting it to produce a pseudorandom bit stream, and then using arithmetic coding to decompress that bit stream based on a statistical model derived from an LLM. This produces text which appears to be sampled randomly from the LLM, while actually encoding the secret message in the specific token choices.”
So using the LLM as a digital filter to get a statistical set of phrases…
Some might remember a simple “coding” technique that made a statistical adjustment to letter frequency to help thwart manual cryptanalysis called a “Straddling Checkerboard”,
https://en.m.wikipedia.org/wiki/Straddling_checkerboard
That originated in the modern form around that “flattens the statistics” around about the late 1930’s. And later became used in the VIC cipher, that outside of “Perfect Secrecy” systems such as the “One Time Pad”(OTP) is the only known “pencil and paper” cipher to defeat the early days of the current US and UK SigInt agencies.
Clive Robinson • February 13, 2025 9:28 AM
@ Bruce, ALL,
Bad news for theft of IP AI merchants.
In Delaware a District Court, judge Stephanos Bibas said by way of a summary judgment the AI usage of copyright material arguments presented by the defendendts “Ross Intelligence”,
“None of Ross’s possible defenses holds water. I reject them all,”
Whilst not yet over or cast in stone, it’s definitely an “Ouch Moment” for those just stuffing everything they can grab into their LLM’s and other current AI systems.
More info at,
https://www.wired.com/story/thomson-reuters-ai-copyright-lawsuit/
Subscribe to comments on this entry
Sidebar photo of Bruce Schneier by Joe MacInnis.
Dancing on thin ice • February 8, 2025 2:30 AM
Anything on the biggest security story going on right now?
An overview of threats would be useful while an attack is in progress.
Those doing oversight or in charge of securing government systems have been fired with no notice.
An unelected billionaire with obvious conflicts of interest has 20 something year olds sending classified data over unsecured email that was accessed from formerly secure government systems.
Normally, many involved all would have trouble passing a security investigation.
Whether anyone agrees politically or not, security concerns should be universal.