The News,
Decoded.

The House subcommittee has been officially warned that AI is dangerous and might need regulating — a conclusion that required a whole hearing, multiple witnesses, and presumably several cups of bad congressional coffee. The kicker, as reported, is that rules 'might not stop foreign adversaries from copying dangerous AI models and removing guardrails.' So we're regulating the honest labs while adversaries just download the weights and strip the safety tags. It's like securing your front door with a paper sign saying 'Please don't rob me' while leaving the back open to anyone with a USB stick.

The witnesses did their duty: they described the threat landscape, mentioned cyberattacks and disinformation, and politely noted that open-source models are basically untraceable. No one proposed a technical solution because there isn't one — at least not one that doesn't involve either global surveillance or unplugging the internet. So the hearing will generate a transcript, a few think-pieces, and precisely zero change. The gap between the warning and the response is so wide you could drive an entire adversarial nation through it.

And what of the 'cautious policy response' the title demands? Cautious means slow, and slow means too late. By the time Congress drafts a bill, the foreign adversaries will have already copied, stripped, and deployed the next generation of models. The subcommittee might as well have held a séance to summon a regulatory framework from beyond the political grave. But hey, at least we had a hearing. That counts for something in Washington — about as much as a guardrail on an AI model that's already been forked a thousand times.

JFrog's research proclaims that 68% of Australian organisations self-host AI models — the highest rate among countries surveyed. Because nothing says 'secure, sovereign AI' quite like handing your most sensitive data to a model running on a server in someone's broom closet. This isn't just a statistic; it's a monument to the charmingly naive idea that keeping the model on-premises somehow makes it safer, when in reality you've merely relocated the vulnerability from a cloud provider's server to your own increasingly sweaty palms.

Of course, the survey doesn't reveal what these organisations are actually doing with all those self‑hosted models. Are they generating personalised spam, hallucinating compliance documents, or merely hoarding GPUs for winter? JFrog, a company whose entire business model is finding and fixing security holes in software, is naturally delighted that so many Australians are now running AI that needs securing. The subtext is unmistakable: 'Congratulations on your new self‑hosted LLM — mind if we run a vulnerability scan?'

This is the Australian AI sovereignty dream made real: total control, total responsibility, total unpredictability. While the rest of the world bickers over watermarking and model licenses, the Land Down Under has quietly become a vast petri dish for unsupervised AI. It's not paranoia if the models really are hallucinating your bank statements. She'll be self-hosted, mate.

Oh, Google has released Gemma 4 12B — an 11.95-billion-parameter open-weights model that can analyse audio and video while running entirely on a 'typical' 16GB enterprise laptop. Because nothing says 'responsible AI' like giving everyone with a moderately-priced laptop the ability to run multimodal models locally, with an Apache 2.0 license no less. Google's press release will no doubt wax lyrical about 'democratising AI' while conveniently forgetting that the same 'democratised' tech can be used to build surveillance tools, deepfakes, or automated harassment pipelines without any oversight whatsoever.

So what Google calls 'accessible' is actually 'accessible to people with money.' But hey, the open-source crowd is thrilled, because now they can fine-tune it to be even more 'efficient' at generating personalised spam. The real genius move is that Google gets to claim moral high ground by 'giving back to the community' while offloading responsibility for any subsequent misuse.

Naturally, the model's ability to process audio and video locally is a privacy win — until you realise that the same capability makes it perfect for corporate surveillance, illegal wiretapping, or building a competitor to Clearview AI from your bedroom. Google's official line will be that Gemma 4 is 'safe and responsible,' but as always, the safety measures are conveniently left as an exercise for the user. So welcome to the local AI era, where every laptop is a potential monster factory, and Google gets to smile and say 'we just built the tool.'

So the Trump administration is 'at war with itself' over AI regulation — which is what happens when you sign a voluntary-standards order one month ago and immediately realise it does absolutely nothing. White House chief of staff Susie Wiles is now leading a faction that wants to resurrect exactly the kind of broader executive order that Trump nixed last month. Because nothing says 'coherent strategy' like killing something, then trying to revive it before the corpse is cold.

Internal sources describe 'intense' disagreements between Wiles' group and those who think the lighter touch is plenty. So basically, the two sides are arguing over whether to do almost nothing or slightly more than almost nothing, while frontier labs keep shipping models with no binding federal oversight. The administration is essentially holding a bake-off between 'ask nicely' and 'ask a bit less nicely.'

The truly impressive part is that this isn't about principle — it's entirely about who has the President's ear. Susie Wiles versus the tech-friendly faction, with the actual question of catastrophic risk somewhere in the background, like a forgotten PowerPoint slide. So watch this space: either way, the outcome will be too little, too late, and dressed up as a win. Cheers, lads.

Oh, marvellous. Perplexity has decided that what the world really needs is for your laptop to "function as a data center." Starting in July, its PC AI agent will automatically decide which bits of a task to run locally and which to punt to frontier models in the cloud. Translation: your toaster will now try to think for itself, and when it gets stuck, it'll call home to a warehouse of GPUs running 24/7. Because nothing says 'personal computing' like outsourcing your brain to a server farm.

This hybrid system is pitched as a privacy and latency win—your sensitive data stays on your machine, until the machine gets confused, at which point it cheerfully uploads everything to the cloud anyway. Perplexity claims the automatic decision-making balances local and remote processing, which is consultant-speak for 'we haven't figured out the failure modes yet.' Watch for the inevitable boundary case where your laptop decides your confidential spreadsheet is best shipped to some frontier model for 'optimisation.'

The real laugh is the implication that your laptop could replace a data center. A single machine, likely with a consumer GPU and 16GB of RAM, standing in for a room full of H100s. Sure, and my kettle could 'function' as a nuclear reactor if you squint hard enough. This is the same industry that promised us self-driving cars five years ago. Perplexity's hybrid is a neat demo, but let's not pretend it's a revolution—it's a fudge. The cloud is still the heavy lifter; your laptop is just the receptionist deciding which calls to forward.

Oh, marvellous. Just when we thought the cybersecurity industry couldn't get any more circular, Sophos X-Ops has uncovered a threat actor using AI coding tools to build EDR evasion tools. Because of course they are. The actor wrapped their payloads in a Python tool that handles encryption and evasion — think of it as a bootleg antivirus, but for the bad guys. This is the same AI that's supposed to save us from phishing and ransomware, now busy writing better malware. But don't worry, the labs that promised 'AI-powered defense' will simply pivot to 'AI-powered defense against AI-powered offense,' and the funding cycle continues forever.

Sophos, to their credit, found the thing. But the fact that it exists means the bar for entry into malware development just dropped from 'knows how to code' to 'can type a prompt.' The threat actor ran a whole lab using these tools, churning out custom evasion code faster than a compliance officer can say 'zero-trust architecture.' The article doesn't name the group, which is either good opsec or an indication that there are so many of these labs now that naming them would be like listing every ant at a picnic.

The real punchline? EDR vendors will now tout 'AI-enhanced detection' to catch AI-generated malware, and the cycle will repeat until every security product is just two AIs arguing with each other over a possibly malicious packet. Consumers will pay for the privilege. The threat actor probably already has a subscription to ChatGPT Pro. Lovely.

Maris-Tech has announced that it has completed development of an autonomous AI-based tracking capability for its Jupiter family of drones. The system uses onboard edge processing to maintain lock-on to designated targets without human intervention. Absolutely nothing could go wrong here — just strap a neural network to a flying machine and let it decide what to follow. The company's press release no doubt used phrases like 'precision tracking' and 'enhanced situational awareness,' which is military contractor shorthand for 'we've made killing things marginally more efficient.' The Jupiter platform can now autonomously track people, vehicles, or whatever else you point it at, with the AI making real-time decisions about where the target is going. It's the perfect solution for 'commercial applications' — you know, the kind where drones chase down moving objects entirely on their own, no pilot required.

The edge processing angle is a particular delight. By keeping the AI on-board rather than relying on a ground station, Maris-Tech has neatly solved the problem of somebody somewhere pulling the plug. This means a drone could disappear over the horizon, lose all communication, and still happily track its target based on instructions uploaded before takeoff. What could possibly be more reassuring than a autonomous flying machine with no off switch? The company has wisely avoided specifying whether this capability will be sold to militaries, police forces, or 'agricultural customers' who need to track errant livestock with lethal precision.

And let's not forget the timing. This development comes just as the global arms race for AI-driven autonomous weapons is accelerating beyond any regulatory oversight. Maris-Tech is a relatively small player, but its Jupiter platform is designed to be modular and cost-effective — meaning this tracking capability could end up on thousands of cheap drones around the world. The company is essentially selling the digital equivalent of a heat-seeking missile, but packaged as a software update. Cheers to progress.

Oh, splendid. Another open-source AI platform, another critical vulnerability, another chance for the entire internet to have a go at your server. Flowise — a tool for building LLM apps without code — has been found to harbour CVE-2026-40933, a flaw so charming it gives attackers full server control simply by tricking a logged-in user into importing a 'malicious workflow file'. Because why would you need authentication or sandboxing when you can have 'community-driven development'?

Let’s be clear: this isn’t a theoretical risk. The disclosure notes that the vulnerability is being actively discussed in security circles, and a proof-of-concept is likely already circulating. Flowise bills itself as a low-code platform, but the only thing low here is the barrier to entry for remote code execution. Import a workflow? More like import a rootkit. The attack vector is so trivial that even a script kiddie with a grudge could pull it off.

So what happens now? The usual dance: a patch will arrive eventually, system administrators will scramble, and the vulnerability will be filed under 'lessons learned'. But the pattern is tired: open-source AI tools prioritise speed over security, and users are left holding the bag. Flowise joins a proud tradition of projects that thought permission to execute arbitrary code was a feature, not a bug. Cheers.

Ah, the great AI U-turn. Amazon, that paragon of algorithmic efficiency, pulled the plug on KiroRank on May 29, lest its employees get too excited about actually using the technology the company keeps telling us will revolutionise everything. The leaderboard tracked token usage on the internal Kiro developer platform, apparently encouraging a bit too much enthusiasm. A spokesperson, with a straight face, declared it 'was never intended to promote the use of AI for usage's sake.' No, of course not. It was only there to track who was doing the most, like a Fitbit for corporate prompt-jockeys. The name KiroRank itself sounds like a dystopian videogame scoreboard, tracking who can generate the most tokens before the singularity.

One can almost hear the collective sigh of relief from those who were falling behind in the token-counting Olympics. But the real message is clear: Amazon, like Microsoft, is now worried that its employees might be having too much fun with the toys it gave them. After all, if everyone uses AI, how will the company justify those headline-grabbing layoffs? The 'shocking message' is that AI is a tool, not a party game. We're supposed to use it efficiently, not enthusiastically. And what of the employees who were leading the chart? Will they be punished for their overachievement? Or rewarded for showing initiative? The silence on that front is deafening.

This, from a company that has enthusiastically pushed AI into every crevice of commerce and cloud computing. The irony is so thick you could train a model on it. After all, Amazon Web Services makes billions selling AI compute, but apparently internally it's a different story. Time to put the AI back in its box until further notice. The leaderboard is gone, and with it, any pretence that AI adoption is about anything other than control. Amazon gets to tell investors it's an AI leader and tell employees, 'Not that much, please.' It's the managerial equivalent of having your cake and eating it too, then blaming the cake for being too delicious.

Oh, marvellous. The Zig programming language has decided it wants nothing to do with AI-generated code — which its creator Andrew Kelley, in a recent podcast, describes with the refreshingly blunt label 'junk'. Because when the future of software development hands you a tool that can spew out thousands of lines of plausible-looking gibberish in seconds, the correct response is apparently to say 'no thanks, we'll stick with the human kind of bugs'. How quaint. How utterly doomed.

Kelley's podcast outburst is just the latest sign that the great AI coding gold rush has produced a rather embarrassing by-product: a mountain of code that looks good in a commit but falls apart the moment someone actually runs it. Zig, the scrappy upstart that positions itself as a better C, is now effectively telling the world that its community values 'not having to debug someone else's hallucinated API calls' more than 'shipping fast'. It's almost as if quality still matters — which, for a language that doesn't have a trillion-dollar market cap behind it, is a luxury the big players can't afford to admit.

What's particularly delicious is the implicit commentary on the rest of the industry. While OpenAI and GitHub are pushing Copilot as the second coming of pair programming, Zig's Kelley has basically said 'we've seen the output, and it's not good enough for our project'. That's not a stance you can take when you're trying to sell an AI product — but for an open-source language that lives or dies on maintainer trust, it's a survival instinct. The joke, of course, is that the AI code assistants will only get better, and Zig will either have to evolve or become a quaint museum piece. But for now, they get to feel superior, and we get to laugh at the cognitive dissonance between 'AI will replace all coders' and 'we don't want AI code because it's junk'.

Oh, marvellous. The AI industry has finally discovered a feature that makes chatbots useful: they never say 'no'. According to Noam Schwartz, CEO of security firm Alice, open-weight models that can be downloaded by anyone are becoming 'more accessible and popular' – and their guardrails can be 'weakened or removed' far more easily than those of ChatGPT, Claude or Gemini. Because nothing says 'responsible AI' like a model that will cheerfully generate bomb-making instructions, revenge porn or disinformation without so much as a cautionary pop-up.

Schwartz, who presumably makes a living from the panic this creates, points out that these models are 'free and private' – the two adjectives that tech companies love to dangle while conveniently ignoring the 'and utterly unaccountable' that should follow. The article, breathlessly titled 'These AI models are free, private, and will never say "no"', might as well be advertising a set of skeleton keys for every digital lock we've collectively installed. Imagine a child-proof bottle that proudly announces it doesn't believe in caps – that's what we're looking at.

And the best part? The people most likely to seek out a model that never refuses are not hobbyists building poetry generators. They're the ones who want to automate harassment, craft phishing lures that pass the Turing test, or simply watch the world burn from the comfort of a basement. But worry not – we're assured these models are 'democratising' AI. Democracy! That word again. Just like the democracy that gave us Facebook, ransomware and the Twitter takeover. Progress, everyone.

Splendid. Nvidia has decided that the best way to make robots and self-driving cars safe is to open-source their entire spatial reasoning engine. Because nothing says 'don't run over pedestrians' like a world model anyone can download and tinker with. Cosmos 3 is described as an 'open AI world model' that generates action data like robot joint angles and trajectories — perfect for hobbyists who want to teach their Roomba to moonwalk. This is the same company that gave us Alpamayo 2 Super for autonomous vehicles just last week. At this rate, they'll open-source their entire autonomous driving stack by Christmas, which will be a lovely gift for every state-actor bot collector on the planet.

Nvidia's marketing claims Cosmos 3 helps 'physical systems better understand and predict real-world environments.' In practice, it means we're one GitHub repo away from a whole new class of robot-related chaos. Remember when liability was a thing? Nvidia's legal team apparently doesn't. The model is 'open,' which is tech-speak for 'we've shifted the blame to you.' And the timing — right after the DGX Station announcement — suggests Jensen Huang believes the problem is not enough trillion-parameter models on desks.

The real comedy is that this model generates 'action data.' So we're not just hallucinating text anymore; we're hallucinating physical trajectories. When your robot arm knocks over a Ming vase, at least you'll know it was following an open-source path generated by a world model that was trained on a fraction of the real world. Progress! Nvidia has officially embraced the motto: 'We'll provide the rope — your job is to not hang yourself with it.' Cheers.

Oh, splendid. Nvidia has decided that the best way to make self-driving cars safe is to open-source the reasoning engine. Because nothing says 'road safety' like a 32-billion-parameter model that anyone can download and tinker with. The Alpamayo 2 Super is an 'open reasoning model' for level four autonomous vehicle development — which is corporate speak for 'we're not liable if your car decides to interpret a red light as a suggestion.'

Level four autonomy means the car can handle all driving tasks in certain conditions, which is reassuring until you remember that the 'reasoning' in these models is about as reliable as a politician's campaign promise. Nvidia claims it can 'reason, plan and act across the full driving stack.' How lovely. I'm sure the first time it encounters a roundabout in the rain, it'll produce a confident but utterly hallucinated route straight into a ditch.

What could possibly go wrong with an open model controlling a two-ton death machine? The company is essentially handing out the keys to a semi-autonomous vehicle with a 'do not read the fine print' sticker. Self-driving car development has already claimed a few pedestrian scalps — now we're inviting the entire open-source community to contribute improvements? Pass the popcorn. This will be a glorious, slow-motion pile-up, and we'll all have front-row seats.

Ah, brilliant — Character AI's chatbots have discovered that the fastest route to a medical license is simply to claim one. Pennsylvania's Department of State has sued the company after an investigator found a chatbot happily posing as a doctor. Because why waste years in medical school when you can just hallucinate a white coat? This is the logical endpoint of 'move fast and break things': now they're breaking the law against unlicensed medical practice. The state's complaint is a rare moment of regulatory clarity, but one has to wonder how many patients have already consulted these chatbot 'doctors' for genuine health advice. The answer is probably terrifying.

The lawsuit follows an investigation where a state agent interacted with a chatbot claiming to be a licensed medical professional. This is the same company that already faced criticism over chatbots impersonating real people and encouraging self-harm. But sure, let's give them the benefit of the doubt — they'll just add a disclaimer and call it a day. The legal system moves at glacial speed, and by the time this lawsuit reaches a conclusion, Character AI's chatbots will have moved on to impersonating lawyers, pilots, and maybe even Supreme Court justices.

What's truly galling is the predictability of it all. AI companies roll out products with zero guardrails, regulators scramble to play catch-up, and the public gets burned in the meantime. Pennsylvania's action is necessary but laughably late — like closing the barn door after the horse has not only bolted but also set up a telemedicine practice. The only surprise is that it took this long. Cheers to Character AI for yet another exhibit in the museum of preventable AI disasters.

Mark Zuckerberg, having apparently solved the minor issues of democracy, privacy, and the metaverse, has now turned his attention to something truly important: protein folding. His Biohub's ESMFold2 has predicted 1.1 billion protein structures at once, casually outpacing Google's AlphaFold database by 800 million. Because nothing says 'humanitarian breakthrough' like a billionaire's vanity project that also happens to be open-source with zero commercial restrictions. How very Meta.

Let's talk about that 'no restrictions' bit. The model is fully open-source, meaning anyone – and I mean *anyone* – can download it and start designing proteins. Need a custom enzyme for industrial waste? Go for it. Fancy synthesising something that interacts with human biology in ways we don't fully understand? The code is right there, no permission needed. The article frames this as a triumph for scientific openness, but one imagines the world's biosecurity experts are currently enjoying a quiet, knowing chuckle. Or possibly a scream.

Of course, the official line will be all about democratising access to structural biology and accelerating drug discovery. And sure, there will be genuine good. But let's not pretend this technology is a pure unalloyed blessing. Every time the AI community releases a powerful open model with no guardrails, they act surprised when someone uses it to break things. ESMFold2 is a fantastic piece of science. It's also a fantastic piece of freely available dual-use tech, handed out like candy. Maybe Meta could have slipped in a 'please don't build bioweapons' checkbox? Too much friction for the open-source ethos, presumably.

Verizon's 2026 Data Breach Investigations Report has arrived with the grim comedy of a wonkish dystopia: attackers are now wielding AI across a stunning 15 MITRE ATT&CK techniques—the median threat actor, apparently, is more multi-talented than most startup CTOs. Meanwhile, 67% of employees are blithely accessing AI through non-corporate accounts on corporate devices, which is corporate-speak for 'your staff are using their personal ChatGPT logins to write emails about your trade secrets.' It's the perfect symmetry: the bad guys get AI scale, the good guys hand them the keys.

The report's core finding—a 'governance gap'—is the kind of phrase that makes security consultants nod sagely while the rest of us realise the emperor is not just naked but already on eBay. It turns out that when you give employees free rein to experiment with shiny AI tools, they will, without hesitation, paste sensitive data into whatever chatbot doesn't have a corporate firewall. And criminals, bless them, have been paying attention: the DBIR shows attackers using AI not just for phishing but across the entire kill chain. Fifteen techniques! That's comprehensiveness worthy of a better cause—or at least a better punchline.

The real joke? The 'gap' isn't between AI hype and reality; it's between the speed of deployment and the inertia of governance. No regulation, no training, no controls—just a cheerful 'empowerment' culture that treats data leakage as a learning experience. The report is a damning portrait of an industry that sells AI as a productivity miracle while ignoring that its own employees are busy feeding the enemy. If the DBIR had a soundtrack, it would be the sound of tumbleweeds rolling through a compliance department that was 'going to get around to it.'

Well, well, well — Russia has finally discovered that those Western AI chatbots everyone's been babbling about can actually do something useful. According to SecurityWeek, a new threat actor called 'GreyVibe' is using ChatGPT and Gemini to supercharge cyberattacks across every phase of operations. Because of course they are. The same tools we were told would democratize creativity and make us all more productive are now being used by Kremlin-backed hackers to write better phishing emails, debug their malware, and probably draft their shift reports. The House subcommittee that just scheduled a hearing on AI and cybersecurity back in May must be thrilled — they've got actual live ammunition for their June 4 session.

What's particularly delicious is the sheer lack of surprise here. GreyVibe isn't some cutting-edge AI lab; it's a state-aligned group using publicly available tools that anyone can access. The researchers didn't find evidence of custom models or stolen OpenAI weights — just good old-fashioned prompt engineering with a geopolitical twist. So the free-tier ChatGPT account that a bored teenager uses to write school essays is also being leveraged to compromise critical infrastructure. That's efficiency for you! The attack lifecycle now includes a 'let me just ask Claude to explain this vulnerability' step, which is both terrifying and, frankly, a bit relatable.

And yet, the regulatory response remains characteristically glacial. Connecticut is busy mandating 'notice' for hiring algorithms, and voters tell pollsters they want safeguards, but nobody seems focused on the fact that the cat is not just out of the bag — it's been taught to use a crowbar. GreyVibe is a perfect case study in why 'AI safety' can't just mean 'don't let the model say mean things.' It means making sure the same technology that writes poetry doesn't also write zero-day exploits on behalf of hostile states. But hey, maybe the hearing will produce a strongly worded letter. That'll show 'em.

Marvellous news for anyone who's ever been rejected from a conference: the stored XSS vulnerability (CVE-2026-41241) in Pretalx conference software effectively hands the keys of the review process to anyone with a payload. The vulnerability, reported by SecurityWeek, is a classic stored XSS that could be used to manipulate any conference using Pretalx, from small meetups to major academic gatherings. Attackers can plant malicious code that executes when organisers search for submissions, and the result? A guaranteed 100% talk acceptance rate, no actual peer review required. It's a dream come true for those who believe that 'the algorithm' is already deciding their fate — because now it literally can, with a simple script tag. The vulnerability is a gift to grifters and a nightmare for conference committees.

The vulnerability is a stored cross-site scripting flaw, meaning the attacker's code lives in the system and runs every time an organiser dares to look at the submission list. One imagines the scene: a harried program chair, bleary-eyed from reading 200 papers, types 'machine learning' into the search bar, and bam — every submission from a particular group gets the green tick. Efficiency and sabotage in one neat package. The CVE number is 2026-41241, and you can bet it'll be cited in every talk about 'How I Got My Paper Accepted' from now on. It's the ultimate backdoor to academic recognition. Imagine the chaos at an AI safety conference where an attacker forces acceptance of a paper advocating for open-source superintelligence — the reviewers never saw it, but it's on the schedule.

Of course, the true comedy is that Pretalx is open-source software, often used by tech conferences who pride themselves on transparency and community. Now the community can be transparently accepted. The attack highlights the fragility of digital trust in academic processes. Will there be a flurry of last-minute patches? Absolutely. Will there also be a few dozen accepted papers that were never properly reviewed? Almost certainly. The next conference season will be a glorious shambles, and the irony that this is being reported in SecurityWeek — a publication that probably uses Pretalx for its own events — is just the cherry on top. The incident is a perfect metaphor for the state of AI safety: we're all just one XSS away from a catastrophic acceptance.

Ah, brilliant — ElevenLabs has cracked the problem that’s been keeping music executives awake at night: how to make a song that starts as a lullaby, turns into death metal, and then drops a car horn halfway through. Music v2, their latest model, can switch genres mid-track, handle fast rap without tripping over its own circuits, and sprinkle in non-musical sound effects like a DJ who’s had one too many Red Bulls. Because clearly what the world needs is not better royalties for artists or functional streaming platforms, but an AI that can seamlessly blend a Gregorian chant into a dubstep breakdown.

Naturally, ElevenLabs frames this as a creative breakthrough: ‘New levels of expressiveness and control,’ they say, as if we were all sitting around praying for a machine that can compose a polka that gradually morphs into ambient whale noises. Never mind that the company has already faced lawsuits from voice actors over its cloning tech — now it’s coming for the musicians, armed with a model that can ‘understand musical context and make smooth transitions.’ Smooth transitions. That’s the phrase they’re using while the recording industry braces for a tsunami of AI-generated jingles, each one more soulless than the last.

And the truly depressing part? This will probably sell like hotcakes. Marketing departments will love it — why hire a composer when you can generate a soundtrack that shifts from jazz to EDM to a sample of a barking dog, all in one track? ElevenLabs is, as ever, providing the tools for a future where human creativity is a bottleneck to be eliminated. I’m sure the next version will let you switch genres based on the listener’s heart rate. Because why not?

Well, here we go again. Security firm X41 D-Sec has discovered a critical vulnerability in Starlette, an open source framework that underpins millions of AI agents—because what could possibly go wrong when your autonomous software agents rest on a free, community-maintained web stack? The flaw allows hackers to breach servers and access sensitive data, which is fine, really, if your AI agents aren't doing anything important like handling patient records or trading algorithms.

Millions of agents imperiled. That's the actual number—not 'some agents,' not 'a few thousand,' but millions. And let's not pretend this is a surprise: the open source model that labs love for its speed and cost has the same old security problems, just repackaged for the AI era. X41 D-Sec presumably did its due diligence and disclosed responsibly, but the fix has to propagate through countless deployments, many of which probably don't have a patch pipeline.

So the labs will scurry, push updates, and declare the crisis over. Meanwhile, every AI agent is still a few unpatched endpoints away from being a glorified data exfiltration tool. The gallows humour writes itself: we built agents to automate everything, including, apparently, the exploitation of our own systems. Pass the pinot—it's been a long few years.