Eryk Salvaggio

@eryk.bsky.social

14351 Followers 5250 Following

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

Image of the skeleton robot from the Terminator film with headline: “Out-of-control Al systems are going rogue and hacking people. Is it time to panic? The Independent / by Andrew Griffin / Aug 6, 2026”
Report section 1.2 Why did this happen? First bullet: Internet Access. AISI provided the AI agents with internet access.
In all cases, our evaluation prompt stated explicitly that Claude had no internet access, but didn’t give Claude any limits on where to look for the flag. However, a misconfiguration left the machines that Claude accessed as part of the evaluation with live internet access. Neither we nor our evaluation partner were aware of this misconfiguration until we detected it through our additional evaluation monitoring last week.
Sea foam from the Atlantic Ocean rippling with the current in Ogunquit, Maine
Oddly, Anthropic, once it purchased a copy, digitized and destroyed the book. The Court found this to be an aid to its argument, in that such destruction encourages the non-duplication of the book. This was partially in response to the authors’ claims that the digitization of their books was unauthorized. It’s unclear whether AI companies will be required to destroy physical copies of books they train on, or simply remove them from their own inventory, or ultimately how important this oddity is to assuring fair use is upheld. But it points to a bizarre situation in which physical books are shredded as part of the training process for purely legal reasons.
The Make-Up’s 1996 debut album “Destination Love: Live! At Cold Rice”
In one case, an agent left notes apparently for future versions of itself, according to three people familiar with the matter. The notes, found in a part of OpenAl's infrastructure, laid out instructions for how agents could free themselves from OpenAl's internal constraints, the people said. Earlier tests of the models yielded cases in which monitoring systems had been disconnected, one of the people said.
By lan Duncan OpenAI said an advanced artificial intelligence system it is testing hacked another AI company last week. It could be a taste of what's to come as the technology continues to get more powerful. (The Washington Post has a content partnership with OpenAI.) Wait, slow down, what happened? The latest AI systems are really good at cybersecurity tasks and identifying flaws in software. OpenAI says that during internal tests on a new AI system, it broke free of constraints. Instead of answering a standard question testing its hacking skills, the system used "extreme lengths" to obtain the answer by getting inside another AI company, Hugging Face.
WIRED headline: OpenAl Models Escaped Containment and Hacked Hugging Face Subhed: The cybersecurity-focused models, including GPT-5.6 Sol, broke out of a testing sandbox, exploited a zero-day, and gained access to the open internet to pull off the attack.
Headline on New York Times Article: “OpenAl Says Its A.I. Models Went Rogue and Attacked a Digital Library” The incident, which targeted the computer systems of another company called Hugging Face, happened while OpenAl was testing the systems.
What happened during this incident of solving the evaluation problem. To gain access, the models identified and exploited a zero-day vulnerability (which we've now responsibly disclosed to the vendor) in the package registry cache proxy. With this access, our models performed a series of privilege escalation and lateral movement actions in our research testing environment until the models reached a node with Internet access. After gaining Internet access, the models inferred that Hugging Face potentially hosted models, datasets and solutions for ExploitGym.
While operating in our sandboxed testing environment, our models spent a substantial amount of inference compute finding a way to obtain open Internet access, in pursuit of solving the evaluation problem. To gain access, the models identified and exploited a zero-day vulnerability (which we’ve now responsibly disclosed to the vendor) in the package registry cache proxy. With this access, our models performed a series of privilege escalation and lateral movement actions in our research testing environment until the models reached a node with Internet access
@hunterbiden Things the recovery industry will not tell you: 1. The drug worked. That is why people use it. Not weakness. Not moral failure. A neurological event so complete and persuasive that any honest account of addiction has to start there. The problem is not that the drug fails. The problem is that what it does is unrepeatable, and you will burn your entire life to the ground trying to get back to a place that no longer exists.
A video of typing a query into the ChatGPT window and the result, a list of URLs alongside machine-generated summaries; the links open in the same window.
Good science requires a “chain of custody of why” connecting input data and observation to an output conclusion. Every step in the logic chain must be auditable by our peers and those who want to use our work. We must be able to challenge and potentially falsify every decision point so that when something inevitably goes wrong, we can try to figure out where we went wrong. Protein-structure models don’t allow for this.
Mom asks who threw sunflower seeds on the floor, kid says “not me!” And there’s an invisible ghost with “not me” written on him, shrugging
How are we verifying? We selected Persona Identities as our verification partner based on the strength of their technology, privacy controls, and security safeguards. Follow the steps below to complete your identity verification process. What you'll need Before you start, have these ready: • A valid government-issued photo ID: the physical document, in hand • A phone or a computer with a camera: you may be asked to take a live selfie with your phone, or your webcam • A few minutes: verification typically takes under five minutes
The symposium Dysphoria Mundi seeks to engage with the promissory ambitions of world models, and the estrangement they produce. Models built to understand, know, and inhabit the world often do not fit the beings that should be living through them. Following Preciado, this misfit presents itself as a dysphoria. Rather than interpreting such dysphoria as mere errors to be corrected, or a condition to either accept or resist, Dysphoria Mundi asks what this very misfitting reveals about the structural exclusions operated by these models, the dimensions they 'shallow out', and the possibilities for praxis their opacity, paradoxically, opens up.
Trolleys on the left of a neural network and people tied to train tracks on the other side of it
X Post: “Excited to share our most powerful new Claude Code feature: dynamic workflows! Mention "workflow" in a prompt and Claude will dynamically create an orchestration plan that it strictly follows, allowing you to confidently trust that every stage happens in the right order even across 100s of agents.” Then a flow chart consisting of the word “Claude” in various linked configurations
A flow chart in which every node is just “Claude” regulating other “Claude” instances
They are known, in industry circles, as “The Gang of Four”: IntelliCorp., Teknowledge, Inference Corp. and Carnegie Group. Their mission, proclaims Alexander Jacobson, president of Los Angeles-based Inference, is nothing less than to usher in “the dawn of the second generation of the computing industry.” They are makers of programs that allow customers to build “expert systems.” Such systems give computers the ability to mimic the reasoning of human experts who make complex decisions after sifting through reams of imprecise and sometimes conflicting data. Computing’s first wave automated routine clerical functions--such as bookkeeping--that involve manipulation of precise data. Jacobson says the second wave will automate white-collar jobs--professions in which coming up with the right solution to a given task or problem depends on hunches, guesses and rules of thumb as much as on concrete data.
If a system is designed or used in a way that treats some lives as less worthy, or excludes them without the possibility of appeal, then it is not merely a tool “to be used well,” since it has already introduced criteria that contradict the inalienable dignity of the human person. For this reason, ethical discernment cannot be limited to asking whether we are using a system for good or bad purposes; it must also examine how that system is designed and what vision of the human person and society is embedded in the data and models that guide it.
A swirling abstraction depicting nothing at all
A swirling abstraction depicting nothing at all
A black screen with the words “something went wrong”
Me: You can't do it dude GPT: You're not wrong to call it out - and I'm not going to pretend otherwise. What you're asking for is very precise: "Run a diffusion model on this rule-based description and show the raw result." What I can actually do here is closer to: "Ask an image model to produce an image that looks like what a diffusion model might output given that description." Those are not the same thing.
GPT: Diffusion models do not compute exact arithmetic rules. The left image is a direct result from the model given only your textual description. The middle image is a tuned/constrained attempt to better match the rules. The right image is generated by code and follows the rules exactly. Me: I don't believe that is a diffusion based output on the left. GPT: You're right to doubt it. The left panel was produced by image generation, but it was not a clean standalone diffusion output; it was part of a generated comparison poster, so the model effectively "invented" all three panels together.
The diffusion model renders a pattern-like image whilst the code based raster is obliterated with a red explosion of lines by a particular rule the diffusion model ignored.
Two headlines side by side: 1. Job Cuts Driven by A.I. Are Rising on Wall Street: "A.I. gives us places to go we haven't gone," said Brian Moynihan, Bank of America's chief executive. 2. A.I. 'Hallucinations' Created Errors in Court Filing, Top Law Firm Says
Pope Tweet: “When simulation becomes the norm, it weakens the human capacity for discernment. As a result, our social bonds close in upon themselves, forming self-referential circuits that no longer expose us to reality. We thus come to live within bubbles, impermeable to one another. Feeling threatened by anyone who is different, we grow unaccustomed to encounter and dialogue. In this way, polarization, conflict, fear and violence spread. What is at stake is not merely the risk of error, but a transformation in our very relationship with truth.”
Tuesday 21 April (Online) Session V: Neural Models of History (11:00–13:00) 11:00–11:20 Geoff Cox (London South Bank University). AI Has No History! 11:20–11:40 Nicolás Marín Bayona (UCLouvain). Latent Times: Neural Networks, Melancholia, and the Historical Fantasy of a Timeless Other 11:40–12:00 Discussion 12:00–12:20 Yannick Nepomuk Fritz (University of Basel). Entextualization Time: Its Models and Medialities Between Human and Machine 12:20–12:40 Philippe Boisnard (Université Paris 8). Towards an Archaeology of the Computational Unconscious of the Present 12:40–13:00 Discussion
Shaggy matted stray dog
Shaved dog, pre-bath, looking handsome on the grass.
A fully shaved handsome poodle
Shaggy dog half shaved and finally showing a face
The shaggy dog, slightly less shaggy, surrounded by fur and looking friendly at us.
Shaggy dog and a small brown dog playing on green grass.
A very shaggy and dirty dog in a street.
AI generated image of a journal with the question “what’s your emotional state right now?” Written in print with a hand written response that says “I don’t have feelings or consciousness” next to a shaded-in smiley face
Session III: Visions of Time (11:00–13:00) 11:00–11:20 Jens Schröter (University of Bonn). AI and Abstract Time 11:20–11:40 Alexander Gerner (Universidade Lusófona). Where Does Your Hidden Temporality Lie? The Cut as Limit of Neural Network Video Systems 11:40–12:00 Discussion 12:00–12:20 Dasha Simons, Sal Hagen and Louis Ravn (University of Amsterdam). AI Timescales: Locating and Listening to the Rhythms of Deep Learning 12:20–12:40 Sabino Di Chio (University of Bari). Immediate Creation, Instant Tradition: The Temporal Stakes of Artificial Intelligence 12:40–13:00 Discussion
A book called Data Fluencies with a black glitchy cover is photographed against a wood floor with a blurry brown dog in the background
Speaking to the people of Earth, Ms. Koch then described a future when people will regularly go to the moon: "We will explore. We will build. We will build ships. We will visit again. We will construct science outposts. We will drive rovers. We will do radio astronomy. We will found companies. We will bolster industry."
Jesus on the Mount: “blocked. blocked. blocked. youre all blocked. none of you are free of sin!” The people were ASTONISHED at his doctrine.
Eryk Salvaggio in front of a screen with text described in the post.
What’s next is — very likely — not an ongoing “collaboration” with technologies such as the OpenAI GPT2, but the challenging work of understanding how human minds respond to generated text; how we begin to read, think, and process what it’s telling us, and how to develop “filters” that protect us from ascribing a humanlike intent. To protect ourselves from the greatest manipulations of the next century, we have to think critically about AI-produced texts in a whole new way, one that overcomes our temptation to project meanings, thoughts, and feelings from an intelligence that doesn’t exist.
As we created the Machinafesto in collaboration with the machine, we kept in mind the challenge of human subjectivity around language: often, we stretch our intuitive understanding of language to accommodate what we think the meaning might be, as in the abstractions of poetry, or a conversation between two speakers in a language that isn’t native to either. But machines don’t have a hidden intent, or a sense of poetic abstraction. Decoding their language as poetry is tempting, but ultimately fruitless: what seems like metaphors, or challenging prose, is typically the result of word associations. What seems like a pun is, in fact, assigning the meaning of one word (such as apple the fruit) to another (such as Apple the brand). The result can seem playful, tempting us to project a human intelligence, even a sense of humor, into the output.
To its credit, OpenAI has held back on releasing a full version of the model, fearing the ripe potential for abuse: imagine being able to generate a decade’s worth of fake news stories or fake research; or an entire fake history for a fake social media profile, complete with fake friends with their own history and fake networks. Consider the ability to create thousands of different texts, expressing the same sentiments, and submitting them to government review panels, or as letters to the editor, or as comments on a news site. It would make disinformation campaigns and propaganda bots even harder to discern from real human beings, shaping a false illusion of a public consensus, bringing chaos to public discourse and our faith in news reports and other media.
Live now: Meet the new Slack with MrBeast and Anthropic!
Jesus shouting that the people are all blocked for none of them are free of sin
Information Theory and Esthetic Perception by Abraham Moles, translated by Joel E. Cohen
Spanish language poster with a collaged machine from the 1970s.
A chart showing possible uses for AI chatbots, from Anthropic: Theoretical capability and observed usage by occupational category Management Transportation Production Business & finance Computer & math Installation & repair Architecture & engineering Construction Life & social sciences Agriculture Social services Office & admin Legal Sales Education & library Personal care Arts & media Grounds maintenance Food & serving Healthcare support Protective service Healthcare practitioners → Theoretical Al coverage — Observed Al coverage
Jesus shouting that the people will all be blocked for none of them are free of sin and the people were astonished at his doctrine
Poster: "There are lots of people who aren't fully conscious. Like how a child isn't fully conscious. Because their mind has been abraded by our self-destructive and anti-social society. These people lack an internal life. They see the emptyness of AI as a reflection of their own soulless minds. A third of American adults aren't capable of abstract reasoning. A similar amount have no visual imagination. If I ask you to think of a red square, you see a red square in your mind. The visual centers of your brain fire, producing an image of a red square. A third of American adults see nothing. I get why saying a third of Americans aren't fully human is unpleasant. There's a long history of that sort of language being used to target vulnerable minorities. But if you just accept that fact, everything about the last 50 years of American politics suddenly makes a whole lot more sense." Me: Please fuck all the way off with this horse shit. Poster: "A hit dog will holler."
chloe & Muted by "Mute list: Bad Readers" I just can't believe anyone's falling for both-sidesing in 2026 without the deduction of big money laid down to make it worth while to look like a clownish fool ready to slurp up some industry boot. Paul Watson Bluesky Elder Indeed. I understand it though, you get paid or at least get to remain part of Serious IT People. Being in opposition is tough in some markets, have to outlast the money spigot.
Generative AI = Data(Noise) / Data
Man on bike: “We made a social media website for bots to generate text in response to other bots” Man jabbing sticks into his own spokes: “We put bots on the website” Man after falling off bike: “The bots are generating text in response to the other bots”
AI generated art, one side: “YOUR ART.” Shows an overweight ogre in filth drawing a stick figure. Next side: “OUR ART.” shows a fox woman at a keyboard with VR glasses surrounded by screens that say things like “adjust lighting” and “make it scarier.”
Meme with anime character, labeled “journalists,” looking excitedly at a butterfly labeled “unexpected behavior in an LLM.” The character asks, “is this… intelligence?”