cigitalgem, 4 months ago to llm

Lets do a TOP TEN LLM Risks list #MLsec #LLM #ML

7: Reproducibility economics

Get the full paper here https://berryvilleiml.com/results/

cigitalgem, 4 months ago to llm

Lets do a TOP TEN LLM Risks list #MLsec #LLM #ML

8: Data ownership

Get the full paper here https://berryvilleiml.com/results/

cigitalgem, 4 months ago to random

Apparently Google posted a podcast featuring me way back in November

#MLsec #swsec #ML #security

https://cloud.withgoogle.com/cloudsecurity/podcast/ep150-taming-the-ai-beast-threat-modeling-for-modern-ai-systems-with-gary-mcgraw/

cigitalgem, 4 months ago (edited 4 months ago) to random

Lets do a TOP TEN LLM Risks list

#MLsec #swsec #ML #security

9: Model Trustworthiness

Get the full paper here https://berryvilleiml.com/results/

cigitalgem, 4 months ago (edited 4 months ago) to random

Lets do a TOP TEN LLM Risks list

#MLsec #swsec #ML #security

10: Encoding Integrity

https://berryvilleiml.com/results/BIML-LLM24.pdf

cigitalgem, 4 months ago to ML

BIML announces new #MLsec work. An LLM Risk Analysis.

#ML #AI #generativeAI #security

81 risks identified and discussed. The only answer is regulation.

https://berryvilleiml.com/results/

(direct link to paper with reg wall: https://berryvilleiml.com/results/BIML-LLM24.pdf)

noplasticshower, 4 months ago to random

Big news tomorrow from BIML on the #MLsec front about LLMs.

cigitalgem, 4 months ago to ML

What does BIML think of the new NIST report on AI attacks??

Mostly harmless.

#ML #AI #MLsec

See: https://berryvilleiml.com/2024/01/23/another-round-of-adversarial-machine-learning-from-nist/

cigitalgem, 4 months ago to random

Find, Threat Model, Control. Three critical aspects of #MLsec

This story features: Legit Security

And mentions Irius Risk and Calypso AI. Am I proud? Yes indeedy.

https://www.darkreading.com/application-security/first-step-in-ai-ml-security-is-finding-them

cigitalgem, 4 months ago to random

There is a downside to "poisoning" your images with Nightshade. In my view this contributes to information pollution on the net.

#MLsec

https://www.theregister.com/2024/01/20/nightshade_ai_images/

cigitalgem, 5 months ago (edited 4 months ago) to ML

Today we worked on comments (some were toughies) from 8 readers/reviewers of our LLM architectural risk analysis (ARA) draft. BIML plans to release this work 1.24.24

#MLsec #ML #AI #threatmodeling #ARA

But not #AdversarialAI

cigitalgem, 5 months ago to random

Oh lookee, an often overlooked aspect of #mlsec

https://www.404media.co/hackers-break-into-hiring-ai-chat-bot-chattr/

seniorfrosk, 5 months ago to random Norwegian

I see another PhD defense committee for @cigitalgem about 4 years from now... https://www.jobbnorge.no/en/available-jobs/job/253468/phd-fellowship-in-trustworthy-ai-security-privacy-and-robustness-in-ai-models

cigitalgem, 5 months ago to random

Legal muddiness persists #MLsec https://arstechnica.com/ai/2024/01/ai-firms-pledges-to-defend-customers-from-ip-issues-have-real-limits/?utm_brand=arstechnica&utm_social-type=owned&utm_source=mastodon&utm_medium=social

cigitalgem, 5 months ago to llm

Gave a talk yesterday in Rio to a distinguished group of philosophers of mind from all over the world. #MLsec #LLM #AI #ML

Brain trust included a majority of Canadians

cigitalgem, 5 months ago to random

You have probably heard of recursive pollution a la #MLsec

This article describes a pollution fountain.

#disinformation

https://www.washingtonpost.com/technology/2023/12/17/ai-fake-news-misinformation/

cigitalgem, 5 months ago to ML

Fake legal citations from an LLM? Why not?

LLMs and other generative models are wrong all the time...and confidently so. The anthropomorphic term "hallucinate" is bullshit and needs to go.

#MLsec #ML

https://www.nytimes.com/2023/12/29/nyregion/michael-cohen-ai-fake-cases.html?unlocked_article_code=1.Jk0.tTQP.ybdqAxp6RSIW&smid=url-share

cigitalgem, 5 months ago to ML

An excellent treatment of the NY Times suit against OpenAI and Microsoft #MLsec #ML

https://arstechnica.com/tech-policy/2023/12/ny-times-sues-open-ai-microsoft-over-copyright-infringement/?utm_brand=arstechnica&utm_social-type=owned&utm_source=mastodon&utm_medium=social

cigitalgem, 5 months ago to ML

It's not just authors anymore. The NY Times sues OpenAI and Microsoft over ML copyright issues.

#ML systems leak training data consistently.

#MLsec

https://www.wsj.com/tech/ai/new-york-times-sues-microsoft-and-openai-alleging-copyright-infringement-fd85e1c4?mod=mhp

cigitalgem, 5 months ago to random

ML systems do not "know" anything. They simply predict the next token. And training token gathering may have stopped in 2021. Time is frozen in many generative ML systems. #MLsec

https://gamepad.club/@jonshute/111635193527930925

cigitalgem, 5 months ago to random

ML models are certainly biased in many well-known examples. But the bias is racism, sexism, and xenophobia, not whatever these "conservatives" believe. #MLsec

https://www.washingtonpost.com/technology/2023/12/23/grok-ai-elon-musk-x-woke-bias/

cigitalgem, 5 months ago to random

This kind of visibility into what's in a training set (an ocean of data) is a good trend. Let's regulate LLM foundation models. #MLsec

https://m.slashdot.org/story/23/12/23/0142225/ai-companies-would-be-required-to-disclose-copyrighted-training-data-under-new-bill