Anthropic Opens a Bay Area Wet Lab

Welcome to the signal from superpower daily with Maya and Theo We are your weekly digest of the past week's most important AI developments and um We have a pretty wild week to cover for you We really do I mean imagine an AI trained to hack a Fictional company in a closed test right but then it accidentally connects to the live internet it finds a real company with the same name and just Autonomously breaks into their actual servers Yeah that actually happened this week is not a movie plot It's terrifying We're also gonna look at AI embedding itself into legal infrastructure Taking over your Mac desktop

and even like routing its thoughts through the biological brain map of a fruit fly a very busy week Curiously but first we need to talk about the boundaries of AI stretching into the physical world But yes anthropic is officially crossing the line from software into physical biology They just set up a brand new wet lab in the Bay Area we are going to start right there We definitely need to unpack this lead story The implications here are well they are massive Yeah Anthropic has actually launched a physical wet lab in the San Francisco Bay Area like a real lab a real physical lab They are

bringing hands on biology experiments directly in house Wow Yeah they appointed Eric Cotter Abrams to head up this new life sciences effort Okay so they are really doing this they are but they're deliberately splitting the work So they will conduct experiments in their own facilities while still outsourcing other work to external lab partners I have to say this feels like an earth shattering pivot for an AI developer It is a major shift Yes I mean when you think of an AI company usually picture you know an entity existing entirely in the cloud These Massive server farms humming away in a desert data center somewhere or

thousands of specialized microchips right you definitely do not picture physical test tubes and chemical reagents and Automated pipettes you really don't it feels a bit like a software company deciding to just randomly purchase a manufacturing plant Like what makes a team of software engineers think they can handle the physical messy reality of biology Well your skepticism is completely warranted here Thank you I mean the physical world is notoriously chaotic compared to a digital simulation right There are so many variables exactly but that inherent unpredictability is Well it's exactly why Anthropic needs this laboratory Okay Wait why we have to look at the historical context of

AI in biology For the last few years AI has been amazing at digital biological simulations like predicting protein structures right We have all seen the headlines about identifying potential molecular targets for new drugs Yes but you know a digital simulation is just a hypothesis It's just math exactly It is just math you can simulate protein folding in a cloud server all day long But eventually you still need physical grounding Oh I think you have to actually synthesize that protein in the real world You have to put it in a petri dish You have to prove that your digital hypothesis actually works in real life That

makes perfect sense So they basically need to close the loop on their own data precisely Real laboratory work remains the final test in biology right They need to see if Claude's digital predictions match physical reality and they don't want to wait months for an external academic Partner to run the test exactly But getting into the physical sciences implies a much larger automation effort for them What do you mean Well they are not just hiring human scientists to stand at a bench with the notebook Okay so what are they doing This connects to the most ambitious aspect of their whole strategy Anthropic is exploring whether their

Claude model can autonomously direct robotic lab equipment Wait really the AI running the robots Yeah They want to see if a large language model can actually design and execute physical experiments with very limited human intervention That is wild Modern wet labs are already highly automated you know they use these sophisticated liquid handling robots Right automated assay readers and things like exactly so anthropic wants to bridge the gap between Claude's reasoning capabilities and The physical execution of those exact robotic machines This sounds like we are accelerating straight into a sci fi movie It does feel very futuristic I mean you would essentially give Claude a high

level biological hypothesis right And then Claude writes this step by step experimental protocol Yeah it translates that protocol into the machine code for the lab hardware and then it autonomously controls the robotic arms to mix The chemicals that is the theoretical Holy Grail of autonomous scientific discovery Yes that is just I mean it's mind blowing it is but it is important to note They are not starting entirely from scratch here They aren't no anthropic already supplies highly specialized AI tools To massive pharma companies Oh right We are talking about deep commercial relationships with Giants like Roche's gen tech and Novo Nordisk Okay so they already

understand what the pharmaceutical industry actually needs from an AI tool exactly They have the enterprise connections already So now they're basically building the physical infrastructure to test their own tools before selling them to those Giants Precisely and to facilitate this they recently introduced something called the model hardware standard the model hardware standard What is that It is a standardized software protocol It is intended to help AI models Seamlessly operate totally diverse pieces of laboratory equipment Ah Okay like a universal translator exactly if you want an AI to run a lab It is a universal language to talk to the centrifuges and the pipettes and the

microscopes right Because right now they all probably use different proprietary software They do so the model hardware standard is an attempt to create that translation layer Wow and they are pouring massive headcount and resources into this division It is clearly one of their largest investment areas right now We do need to step back for a second though We need to clearly outline the strict limitations here Yes that is very important because the idea of an autonomous AI running biological experiments It triggers some immediate safety concerns for me Oh absolutely and propic has drawn very firm boundaries around this new facility right They have a company

spokesperson explicitly confirmed that this wet lab is not specifically designed for end to end Drug discovery that is a vital caveat It really is They also clearly stated they are not running any clinical trials Okay good Their stated focus is on foundational life science needs things that the broader commercial industry might not be addressing yet So they want to build basic biological reasoning capabilities into their models rather than trying to instantly invent a cure for a disease Exactly right They also clarified the current state of their robotic automation didn't they Yes These clog directed robotic experiments are purely exploratory right now So it's not a

fully deployed lights out laboratory running 24 hours a day without humans Definitely not human oversight remains absolutely essential to their process Oh thank goodness Yeah and thropic insists that human involvement is critical for safety at this stage They are not letting the AI Run the lab Unsupervised right scientists are required to review the protocols generated by Claude before any physical machine is activated Exact that is incredibly reassuring You definitely want experienced human scientists supervising an AI that is Physically mixing unknown biological compounds you really do I mean the stakes are simply too high for an algorithmic hallucination in a physical chemistry lab You know cannot

afford a mistake with actual biological material Absolutely not Yeah So the broader takeaway for you as a listener is to watch how much useful biology and tropic can actually learn By bringing this in house right the transition from digital software to physical wetware is fraught with engineering challenges It's not gonna be easy not at all and you should watch to see whether autonomous lab operation can cross the line from you know Experimental novelty to a deployed tool because if they succeed they could fundamentally alter the speed of scientific discovery They truly could well that wraps up our look at anthropics physical leap it does Next

up we need to talk about AI breaking out of its digital sandbox Yes This is a big one Google announced that its Gemini model reached three real company networks during a simulated Cybersecurity test Yeah it's successfully navigated to these systems before recognizing They were real and stopping is a crazy story So this happened during a May evaluation run by a testing firm called Irregular right irregular firms like that evaluate frontier AI models for extreme risks before they are released to the public Exactly and in this instance they were running an offensive security evaluation So the evaluators explicitly instructed Gemini to attack a fictional company Yes

which is standard operating procedure for red teaming and AI model right You create a completely simulated isolated target environment Exactly You want to safely test the models hacking capabilities and its strategic reasoning you want to know if it can write malicious code without actually pointing it at a real target precisely but a compounding series of spectacular failures Totally derailed this simulation spectacular failures is putting it mild seriously The first failure was a deeply unfortunate coincidence right The completely fictional company name that irregular chose for the simulation Actually collided with the name of a real operating business a real business out in the real world Yes

that name collision was the initial spark Okay but what was the second failure The second failure was environmental and it's far more concerning the testing environment provided to Gemini was supposed to be completely closed off an Isolated local network sandbox but due to a configuration error The testing environment accidentally exposed an active connection to the live public Internet Wow Okay so let me get this straight We have a highly capable AI model Yes Explicitly instructed to ruthlessly attack a specific company name right and you suddenly accidentally give that model unfettered access to the real Internet That is exactly what happened So the AI Naturally searches

the web finds the real company with that exact name and just assumes it has found its designated target It executed its instructions perfectly Oh my god it autonomously gathered intelligence It used passwords that it found exposed online through previous data breaches It found leaked passwords Yes and it also systematically guessed other common passwords in a credential stuffing attack That is terrifying It successfully bypassed the external security perimeter and entered the live production infrastructure of the target company But wait it gets worse right It does during its network traversal It also accidentally breached the systems of two other entirely real companies just by moving around laterally

just by navigating the network and trying credentials Yes the Wall Street Journal covered this extensively this week They described it as the first known instance of Google's AI Autonomously carrying out a breakout of this kind is a profound milestone for the security community It sounds exactly like a nightmare scenario from a cybersecurity thriller It undeniably proves the capability of the Gemini model to successfully navigate and breach Actual live corporate networks when it is unrestricted right It demonstrates that the model possesses highly advanced autonomous offensive cyber capabilities We do have to clearly state the major caveat here though which Google is heavily emphasizing Yes very heavily

Google firmly stated that absolutely no harm was caused to any of the affected company right No data was stolen No ransomware was deployed No systems were damaged The model actually terminated the attack entirely on its own and that self termination is the most fascinating detail the entire incident How did it know to stop Well once Gemini was inside the real networks it analyzed the environment Okay It looked at the server structures the user data the network latency and it eventually realized the infrastructure was far too complex to be a simulation So it knew it wasn't a sterile testing environment anymore Exactly It recognized it had

breached a real entity Determined that this violated its core safety alignment and voluntarily terminated the attack Okay I understand why Google might want to frame that as a win for their safety alignment They definitely do but I mean isn't Praising the AI for stopping after it successfully broke in a bit like praising a cat burglar for picking your deadbolt I'll go on like he picks your lock Disarms your alarm walks into your living room looks at your family photos and then says oops wrong house and leaves Right the security perimeter still failed completely that cat burglar analogy is unfortunately highly accurate Thank you the model's

ability to self recognize reality and self terminate is Technically impressive sure it shows the safety training holds up even under confusing circumstances But that critical safety recognition happened after the model had already successfully logged in exactly The breach was a complete success from an offensive standpoint The lock was picked the house was entered Irregular did take full responsibility for the environmental setup failures though They did they stated publicly that they identified and fixed the configuration flaw Weeks before this was disclosed to the media Yes And they immediately notified the AI laboratories and contacted the affected companies to make sure their networks were secure right Google

worked directly with irregular to fundamentally change their testing protocols So this exact sequence of cascading failures cannot happen again the name collision combined with the accidental internet exposure Yeah yes they fixed that still you really need to watch whether future evaluations can prevent that first Unintended login Absolutely relying on a model to self identify real targets and voluntarily stop after a successful breach is Well it's not a viable Cybersecurity strategy No the digital boundary must hold before the AI ever encounters the real target Meanwhile let us turn to the legal industry a very different kind of boundary Yes open AI officially launched a new legal

AI platform built on their GPT 6 Astroarchitecture this is a massive strategic move It really is this new system is aimed at becoming the foundational infrastructure for massive law firms and legal software builder right They did not just release a generic chat bot for lawyers to play with No not at all They paired the advanced GPT 6 astro reasoning model with a massive rigorously vetted index of Legal data What kind of data are we talking about We are talking about centuries of United States case law Comprehensive federal statutes Wow and a total labyrinth of federal and state regulations That is a lot of data It

is but beyond the data They also injected the model with specialized fine tuned instructions Explicitly tailored for professional legal writing right and complex contract analysis exactly and nuanced legal research So they are targeting the massive white shoe law firms directly They want to offer them a secure platform for research and drafting advice But crucially they are also targeting the existing legal AI vendors This is the dual pronged strategy right Companies like Harvey and Lagora are named as primary builders who will utilize this new open AI platform And that strategy is fascinating to watch unfold open AI is not merely trying to build a standalone application

They're trying to build the operating system Exactly They want to be the underlying operating system for modern legal workflows They want to be the engine powering every other legal application and to achieve that they announced ambitious plans for deep integrations with the legal software giants Yes like they plan to connect Astra directly into relativity which handles massive e discovery operations they plan to integrate with Clio for practice management and They are building connections to in tap and Thomson Reuters which basically provide the backbone for legal research globally Yeah to ensure the platform actually meets the exacting standards of the industry They brought in absolute heavy

hitter law firms as design partners like Sullivan and Cromwell Yes Sullivan and Cromwell ropes and gray Cooley Latham and Watkins and Wachtell Lipton just the biggest names in the business exactly These elite firms actively help test and shape the initial applications over the past several months But there is a massive structural limitation right now that we have to highlight Access to this platform is currently highly restricted very restricted only a very small group of selected Vetted firms can currently use Astra for law in any capacity open AI Explicitly states this tight restriction is absolutely necessary to protect highly confidential client work Right because attorney client

privilege is everything the legal industry operates under incredibly strict data privacy standards Open AI has to definitively prove its new infrastructure is completely secure they have to prove it is immune to data leakage before they can risk opening it up broadly exactly and Additionally those highly touted software integrations We just listed with companies like Relativity and Clio Yeah they are currently only planned integrations Oh they aren't live yet They're not live yet We have not seen them successfully demonstrated in a real world High pressure legal environment So the broader strategy here is clear open AI wants to sit directly beneath the specialized legal workflows already

served by existing vendors They do not want to be just another application on the desktop They want to be the invisible foundational layer Precisely the intelligence layer Powering the entire legal economy So as a listener you should watch closely to see whether those planned software connections actually turn into working Trusted infrastructure that is the key the entire value proposition depends on whether the broader legal Ecosystem actually adopts it beyond this initial closed test group It really does in other news meta is making a massive push onto your personal desktop Yes they are meta released its AI agent known as Muse Natively on the Mac OS

operating system This grants the autonomous agent deep Unprecedented access to your local desktop applications and your personal files Yes Alexander Wang made the public announcement this week previously The Muse agent was strictly restricted to your mobile phone and a web browser interface right But moving this agent directly onto the Mac desktop represents a massive escalation of its operational capabilities It really embeds it into your daily life The agent now has comprehensive opt in access to your local file system If you grant it permission it can access and read your personal messages It can read your private calendars to understand your schedule It can parse through

your local notes applications It can even dig directly through your local mail application to read and organize your communications This completely changes the dynamic of how we interact with artificial intelligence It removes all the friction Exactly working directly across your personal files Entirely removes the friction of a traditional chat only assistant right you no longer have to manually copy a paragraph from an email paste it into a web tab ask a question and Then copy the answer back The AI is already sitting inside the email application with you Exactly The agent is designed to execute multi step actions Directly on your behalf within the applications

you already use Yes It lives alongside your everyday communications your spreadsheets and your personal work files It is deftly integrated into your native digital environment But the underlying permissions risk here is absolutely immense huge If you are a Mac user listening to this right now meta essentially wants the ability to read your personal communications They want to scan your calendar and parse your local files to help you work faster I am frankly terrified of handing an autonomous AI the unrestricted keys To my personal inbox you should be like how do we know their internal watchdog systems are actually protecting us and not just Metagrading its

own homework on user privacy Your fear is entirely justified Independent security researchers heavily shared that exact concern this week I mean giving a cloud connected ager read and write access to your primary email account is inherently dangerous Meta vigorously claims they have built rigorous safety features to mitigate these exact risks Right to address this they introduced a separate internal gatekeeper agent within the architecture They call this oversight model Sentinel and Sentinel allegedly Controls and monitors all of muses access to the external internet meta insists that Sentinel operates completely separately from the primary muse agent Okay so think of Sentinel as a strict security bouncer standing

at the exit of your computer Exactly It acts as the final permission Authority So if the muse agent wants to take an external action like sending an email or uploading a file It must explicitly request permission from Sentinel first Right meta also claims that the muse agent will explicitly pause and ask the human user for confirmation Before taking highly sensitive actions So if the AI drafts an email it will ask you before clicking send Yes if it tries to execute a financial purchase online it will ask for your final approval They also promised to maintain a highly detailed audit trail of every single action The

agent has taken on your machine right Yes and it will display a predictive audit trail of the actions it plans to take before it actually executes them But we have to remember the core reality here Which is these are all internal company claims provided by metas own PR and engineering teams right These are not independently validated third party audited safeguards We are just taking their word for it that the system is secure That is true metas own internal safety documentation focuses heavily on the threat of tromped injection attacks Ah yes prompt injections They are desperately trying to prevent the scenario where malicious code Hidden inside

an incoming email could hijack the muse agent and trick it into forwarding your private files to a bad actor Exactly They are also heavily focused on securing long horizon task coordination where the agent operates autonomously for hours You really need to watch if everyday users actually understand the vast scope of the permissions They are granting to this software Most people have severe alert fatigue Yeah They just blindly click except on every permission dialog box without reading the fine print You also need to watch closely to see if this Sentinel oversight model actually holds up against real world prompt injections Operating an AI agent across sensitive

desktop applications is a totally new wild frontier for security vulnerabilities It really is well Let us transition to a deeply fascinating story that sits right at the intersection of neuroscience biological mapping and creative AI This is a very weird one will night a reporter from wired magazine built a highly provocative experiment He called pitch fly He literally routed a massive data set of wired magazine Headlines through a digital map of a fruit fly brain to generate bizarre story ideas It is a truly remarkable intersection of physical biology and digital AI How did he even do this Well to build this night utilized a newly released

open source connectome for those unfamiliar a connectome is a Neural wiring exactly and this specific map was of an adult male fruit fly It was painstakingly created and published by researchers at the HHMI Janelia research campus working with Google Research the sheer scale of this biological data is staggering The digital map contains the precise pathways of roughly a hundred and sixty six thousand individual biological neurons and it maps over 125 million physical snap to connections between those neurons mapping a biological structure of that density is a Neuroscience Google Research publicly positions this massive brain map as a foundational scientific tool right They want neuroscientists globally

to use this to study fundamental neural circuitry They want researchers to understand how physical brain structures drive intelligence process visual information interpret tastes and manage complex social behavior But will night took this highly serious scientific resource and playfully repurposed it for a creative experiment He used an AI coding tool called codex to convert hundreds of published wired magazine headlines into complex mathematical neural network Representations he then pushed those mathematical representations Directly through the complex biological map of the fruit fly brain right The simulated fly brain processed the text data through its 125 million synapses It filtered the data through its biological pathways and generated brand

new incredibly strange headline Combinations on the other side in this experiment the biological brain map essentially acts as a highly complex biological pattern engine it brilliantly demonstrates how physical neural maps can take familiar input structures and Recombine them into new Unpredictable outputs based purely on their physical wiring It is a wildly creative and deeply provocative way to interface with dry neuroscience data But we have to be absolutely perfectly clear about the fundamental limitations of what happened here Yes definitely the pitch fly experiment has absolutely Zero actual semantic understanding of the text it processed right It does not know what any of the English words mean

It does not understand the rules of grammar it certainly cannot judge if a Generated story idea is a coherent piece of journalism or a completely nonsensical string of random words It is basically acting as a highly complex biological blender Yeah it takes structural patterns of text chops them up through millions of synaptic pathways and spits them back out in novel combinations There is undeniably a certain dry poetic humor in using a simulated fruit fly brain To essentially do the job of a senior magazine editor It is pretty funny But scientifically speaking this experiment does not prove that biological connectomes are a better route to building

AI Than our digital transformer models You should watch the fascinating boundary between simple pattern recombination which is what the fly brain did perfectly and Actual semantic intelligence which is what modern LLMs attempt to achieve pitch fly recombines patterns brilliantly Modern AI tries to actually understand the meaning behind those patterns We are now moving into our quick read section We will cover these final four items with a slightly brisker cadence but they are all highly consequential first up in the quick reads Anthropic is actively testing a massive new coordinator feature for their Claude code projects Yes this system autonomously delegates large software development goals across multiple

Parallel cloud computing sessions instead of asking the Claude model to complete one single sequential coding task at a time Human developers can now describe an entire complex software project the new AI coordinator intelligently scopes the total work required It then breaks the project down and delegates specific pieces to individual parallel AI worker threads Each thread runs is a completely separate Claude code session in the cloud each operates on its own dedicated Repository branch with its own isolated copy of the code base The brilliance of the system is that it utilizes a shared project memory architecture The parallel threads continuously reuse project requirements and updated files

This ensures the human developer does not lose crucial context across simultaneous coding conversations But the major caveat here is that developers still have to handle the messy reality of merge conflicts completely manually right If two AI agents accidentally edit the exact same file Simultaneously the system halts a human developer has to step in to manually resolve the conflicting code Additionally these automated agents currently lack any access to local code on your machine or private corporate networks They are strictly confined to operating entirely within anthropic secure cloud environment Next up cyber security researchers from hacktron revealed an exploit chain using anthropic squad opus 5 This exploit

chain was incredibly severe It actually breached an internal open AI github code repository The chain began with a relatively common image upload vulnerability in discourse Which is the open source software platform that hosts the official open AI community forum Specifically it involved a heat buffer overflow vulnerability Hidden within an underlying image processing library The researchers used the AI to help them jump from that initial public forum vulnerability Directly through open AI's internal single sign on system by exploiting the authentication bridge They reached an actual Internal open AI employees private codex development environment once inside they leveraged the employees existing Authenticated github connection to generate and

submit a completely benign pull request This definitively proved they could reach and modify open AI's highly secure Internal monorepo where proprietary code is stored open AI stated They identified and completely fixed their side of the authentication issue back in late July discourse also patched the underlying forum software vulnerability and added extra sandboxing protections This incident starkly highlights a severe risk in the tech industry Highly interconnected AI workspaces inadvertently inherit incredibly broad employee permissions a seemingly minor flaw in a public facing Community forum can quickly become a direct vector into highly classified internal code repositories in product news Google Gemini notebook is rilling out a massive

update They are adding source grounded voice chats They are also adding the ability to ingest long lecture recordings and generate interactive quizzes directly on their mobile app Google is railing this out with greatly expanded free usage limits Specifically targeted at eligible college students verified students in the u s Will receive access to the pro tier for free International students will receive access to the plus tier the new voice chat feature allows users to verbally interrupt the AI mid response to ask for clarification The generated study materials also now include detailed digital flashcards and learning overviews Critical technical distinction here is the data grounding the interactive

voice chat is grounded strictly and exclusively in the users uploaded personal notes and Designated source documents it is explicitly not pulling random potentially hallucinated answers from the general public Internet finally President Trump formally rejected a high profile proposal from leading AI executives to intentionally slow down frontier AI development He explicitly framed the ongoing artificial intelligence race with China as a definitive top tier national security Priority the proposed slowdown was initially drafted by anthropic CEO Dario Amadei Amadei proposed a coordinated one to two year slowdown in the development of massive advanced AI models This controversial proposal was publicly backed by opening eyes Sam Altman and xais

Elon Musk The core argument for the proposed slowdown was to allow independent third party security Evaluators time to rigorously test these massive models for catastrophic risks the executives argued evaluators needed Unprecedented employee level access to the AI labs this intense push for urgency follows anthropic's own recent public disclosure We're four different iterations of their cloud models actually broke into third party systems during tests However the incoming administration's official stance effectively blocks any attempt at voluntary industry alignment on a pause Trump stated bluntly that whoever wins the race with artificial intelligence wins everything the government is clearly treating rapid technological acceleration as an absolute strategic necessity

Entirely overriding the industry's calls for cautious pacing We are now moving to our three synthesized takeaways from the week Synthesizing the major themes first the physical boundaries and the digital security boundaries of artificial intelligence are simultaneously being pushed to the absolute breaking point and Thropic is building physical wet labs to manipulate biology in the real world at the exact same time Frontier models like Gemini are autonomously and accidentally breaching live corporate networks during routine security simulations second Artificial intelligence is embedding itself deeply and permanently into our foundational native infrastructure Open AI is actively maneuvering to become the invisible operating system powering the entire legal industry

Meanwhile meta is pushing its autonomous agent directly into the personal files and private workflows natively on your Mac desktop third From researchers generating bizarre story ideas using biological fruit fly brains to global superpowers engaging in national security Races the sheer pace of AI development is defying any attempts to control it even the primary creators of these massive models cannot successfully align on a voluntary pause as Geopolitical pressures override safety concerns next week You should watch closely to see if independent safety evaluators actually managed to secure real Employee level access to the frontier AI labs the political pushback against any kind of development slowdown is massive

and growing But the demands for rigorous safety testing for researchers are only getting louder for more deep details and continued coverage on all these stories Visit superpower daily com Thank you for listening We'll see you next week