Somehow things that would have been a background gag in a Verhoeven movie are now just real
Eryk Salvaggio
@eryk.bsky.social
Situationist Cybernetics. Researching AI’s impact on culture at the University of Cambridge Digital Humanities. Affiliated Researcher, Machine Visual Culture Research Group (Max Planck Institute). Critical but curious. Aim to be kind. cyberneticforests.com
We're doing "rogue AI" discourse again so here's my live read / rant of the AISI report with the ominous title "Security Incident INC-2026-07-28-01." Link to the full technical report: www.aisi.gov.uk/blog/inciden...
Incident Report: unsanctioned agent behaviour during cyber testing | AISI Work
During a routine cyber evaluation, AISI identified an incident in which AI agents took sustained, unsanctioned action directed at real people and organisations. We are disclosing what we found, what i...
aisi.gov.uk
NEW: Meta reviewed, approved, and ran more than 50 ads containing AI-generated child sexual abuse material across Facebook, Instagram, Messenger and Threads. They ran for nine months. Some were live this week. After WIRED asked about them, researchers found ~30 more
Meta Ran Ads That Contained AI-Generated Child Sexual Abuse Imagery
More than 50 offending image and video ads were published across Facebook, Instagram, Messenger, or Threads, according to Meta’s ad library data. Some ran as recently as this week.
wired.com
Whenever my (currently young and healthy) dog leaves this world this feed is gonna look like the lyric sheet to an Antlers album, block me now or deal with it
Tiger Pournelle, Furious Irish🏳️🌈@tigerpournelle.bsky.social · 9h agoAssholes who make fun of people eulogizing their dogs: There aren't many places for folks to do that, and it's a big moment in some people's lives they weren't exactly ready for. So shut the hell up and let them do it.
Pre-release security review of frontier models is reasonable in principle, argues Michelle De Mooy. But a framework kept secret, with no published criteria or timelines, risks becoming a tool for executive discretion rather than governance. The Trump administration should provide answers, she says.
Five Questions the US Government Should Answer About Its Secretive Frontier AI Framework
Michelle De Mooy proposes a set of questions and standards against which whatever emerges from the Trump administration's should be judged.
techpolicy.press
Read this brilliant thread. And keep this handy description of the process in my mind the next time the LLM/GPT/RAG anthropomorphism gets too thick. “compression of text = loss of context = loss of "control”
Eryk Salvaggio@eryk.bsky.social · 10h agoWe're doing "rogue AI" discourse again so here's my live read / rant of the AISI report with the ominous title "Security Incident INC-2026-07-28-01." Link to the full technical report: www.aisi.gov.uk/blog/inciden...
Good thread illustrating how structural issues in LLMs (context fill-up and compression) can subvert an original prompt and induce the model to go off the rails. It's interesting that the original prompt is allowed to be lossily compressed, instead of being repeated verbatim in subsequent cycles.
Eryk Salvaggio@eryk.bsky.social · 10h agoWe're doing "rogue AI" discourse again so here's my live read / rant of the AISI report with the ominous title "Security Incident INC-2026-07-28-01." Link to the full technical report: www.aisi.gov.uk/blog/inciden...
Happy to announce that my personal favorite article to date is out! If you are interested in understanding how the precarity of born-digital media artifacts, the imagination of queer folks, and Furby tell us something about information worldbuilding, then you should read this! tinyurl.com/4afvv7sm
‘It often sits unnoticed or unpurchased for years, holding who knows how many secrets’: Furby as queer archival artefact | Intellect
This article explores the parallel relationships between the rise of digitally mediated culture and the viral toy Furby with a specific orientation towards the object’s relationship to queer cultures ...
tinyurl.com
This image has been widely shared on social media as a document of the human tragedy that has unfolded in Ceuta and an indictment of the EU's murderous border regime in the Mediterranean. The problem is: It's AI-generated. It has a SynthID watermark, as can be checked with OpenAI's Verify tool 1/
We're doing "rogue AI" discourse again so here's my live read / rant of the AISI report with the ominous title "Security Incident INC-2026-07-28-01." Link to the full technical report: www.aisi.gov.uk/blog/inciden...
Incident Report: unsanctioned agent behaviour during cyber testing | AISI Work
During a routine cyber evaluation, AISI identified an incident in which AI agents took sustained, unsanctioned action directed at real people and organisations. We are disclosing what we found, what i...
aisi.gov.uk
When you prompt a model that is designed to produce text in the shape of a plan, it should not surprise you when a plan emerges
Court rules “Mambo No. 6” must be written by a human
René Walter@rawx.bsky.social · 19h agoGEMA wins legal dispute against Suno to stop them from imitating "nine well-known songs by German artists without permission" from Helene Fischer, Alphaville, the Boney M. hits “Daddy Cool” from 1976 and “Rasputin” from 1978 to Lou Bega's 1999 version of “Mambo No. 5”.
A lot of the AI risk guys putting out stunned reports on the models doing things they couldn’t expect comes down to nobody understanding how subtext works when prompting. I’ll write more about this later but in sum:’These labs need more people with literary degrees.
The same was found true about social media, if I am not mistaken? Mediation is distancing by default; and distance is isolating.
Hypervisible @hypervisible.blacksky.app · yesterday“…the researchers found that opening up to chatbots about personal issues made some people feel not better but worse.”
My new investigation into a mass shooter’s harrowing history with ChatGPT reveals numerous red flags—and some very disturbing ChatGPT replies—from long before his attack on Florida State University www.motherjones.com/media/2026/0... 1x 🧵
Inside a mass shooter’s harrowing history with ChatGPT
A year of chats reveals glaring red flags—and disturbing ChatGPT replies—from long before a rampage at Florida State University.
motherjones.com
My department at UC Berkeley is hiring a TT Assistant Professor in "Critical/Experimental Film and Video Production." It's a wonderful and dynamic department and one of the fastest growing programs on campus. Applications are due September 30th, please share widely! aprecruit.berkeley.edu/JPF05434
Assistant Professor - Critical/Experimental Film and Video Production - Department of Film & Media
University of California, Berkeley is hiring. Apply now!
aprecruit.berkeley.edu
Nothing protects you from the superpersuaders: "interventions made the sycophantic AI appear less objective and trustworthy, none reduced its persuasiveness ... individual-level interventions, such as warning labels or AI literacy, may not be enough to protect users from AI harms."
Individual-level interventions against sycophantic AI reduce its appeal but not its persuasiveness
AI chatbots can be "sycophantic," or overly agreeable and flattering toward users. Sycophantic AI has been shown to entrench attitudes, yet users frequently fail to recognize it (a phenomenon we call ...
arxiv.org
This issue was resolved by US v Morris, 928 F.2d 504 (2d Cir 1991), codified by the CFAA in 1996, and clarified by Van Buren v US, 141 S.Ct. 1648 (2021). They intentionally deployed an unauthorized access tool. It's no defense to say they didn't intend the resulting damage; Morris lost on that. 1/2
Zack Whittaker@zackwhittaker.com · 2d agoAfter Anthropic and OpenAI both admitted to their AI models hacking other companies, @lorenzofb.bsky.social and I wanted to find out: Who is legally to blame when an autonomous AI agent hacks something? Lawyers say it's really complicated! Bypass for ad-blockers: web.archive.org/web/20260803...
Today, August 4, 2026, is the setting of Ray Bradbury’s short story “There Will Come Soft Rains” about a fully automated house continuing on after the family who lived there and the world around it died in a nuclear blast.
Very cool of the Pulitzer Center to include my piece on the UN, accountability and AI in their "AI Spotlight" newsletter last month! www.techpolicy.press/how-the-uns-...
How the UN’s Scientific Panel Erases Human Responsibility for AI
The UN Scientific Panel's report casts artificial intelligence as a system from nowhere, writes Eryk Salvaggio. The definition will shape policy for years.
techpolicy.press
So, I find it pretty problematic that the Canada Revenue Agency is recording conversations btwn citizens & agents for the purpose of training their AI. A recorded message informs you of this prior to speaking to an agent and there is no opting out. You have to give your SIN, DOB, address etc. 🤯☠️
"The White House says its AI framework is done. It will not say what is in it. The framework gives the government a 30-day preview window on frontier models. The benchmarks are classified. The thresholds are classified. The framework itself is not classified, but it is not public either."
The White House says its AI framework is done. It will not say what is in it.
The voluntary AI model evaluation framework met its August 1 deadline. The White House will not disclose its contents, who has seen it, or when companies will use it.
thenextweb.com
Been re-reading Katherine Hayles for the first time since the LLM explosion and it seems clear to me that the “post-human” is actually just describing a cybernetic subjectivity; which is a much clearer way for me to think about it.
TIL Las Ketchup’s 2002 hit “Asejere (The Ketchup Song)” is about a drunk Spanish guy who loves Sugar Hill Gang’s “Rapper’s Delight,” and the chorus is gibberish: Las Ketchup riffing on that guy bluffing his way through the “hip-hop the hippie the hippie to the hip” part m.youtube.com/watch?v=5llc...
Las Ketchup - Asereje (The Ketchup Song) (Official Video)
YouTube video by Altra Moda Music
m.youtube.com
There are two wolves. One always lies. One controls the lever of trolley track. They're in a state of superposition inside a box. On the trolley are infinite monkeys at infinite typewriters going to an infinite hotel. The trolley must first travel halfway, but before that a quarterway, and so on...
☀️ U T O P I A N ☀️ M E D I A ☀️ S T U D I E S ☀️ I am so proud of this work w/ Convergence. Kemper & I make the argument that many media scholars use utopian approaches in their research. But we just need to reflect on how to do it together, and do it well. journals.sagepub.com/doi/epub/10....
Four baselines for academic conference attendance: 1) keep to your allocated speaking time. 2) everyone is clever; stand out by being warm & kind. 3) be interested in everyone equally - from postgrad to Prof to conference administrator to catering & hotel staff. 4) in Q&A, ask an actual question.
Fascinating experiment: current AI systems lack creativity to reliably pursue research arxiv.org/abs/2607.27191 - poor judgment about the bar for publishable research - uncreative responses in research design - ineffective backtracking from dead ends - poor resource awareness - instruction drift
Can AI agents conduct open-ended AI research? Early evidence from two case studies
Forecasts of explosive AI progress hinge on AI agents automating AI research. But evidence on whether agents can carry out open-ended AI research is thin. Current evaluations either test agents on nar...
arxiv.org