🧬 DeepMind launches AlphaGenome to spot genetic drivers of disease ↗
DeepMind unveiled AlphaGenome, an AI system aimed at predicting how DNA mutations change gene regulation - basically when genes switch on, where, and how loudly. It can scan huge stretches of DNA in one go, including the sprawling non-coding regions that often feel like biological dark matter.
The pitch is pretty direct: faster identification of which mutations truly matter for things like cancer risk and complex inherited diseases. If it works as advertised, researchers spend less time guessing and more time testing the right things - which sounds obvious, but is kind of the whole game.
🧑💼 Artificial intelligence will cost jobs, admits Liz Kendall ↗
The UK’s technology secretary openly said AI adoption will cause job losses - not the usual “it will be fine, promise” vibe. She pointed to anxiety around graduate entry roles in areas like law and finance, and didn’t pretend there’s a neat number anyone can offer.
At the same time, the government is leaning hard into adaptation: a major push to train millions of workers in basic AI skills, aiming to make the UK a faster AI adopter. It’s the classic tension - yes jobs go, yes jobs appear, no it won’t feel smooth in the middle.
🗞️ UK pushes Google to allow sites to opt out of AI Overviews ↗
UK competition regulators proposed changes that would let publishers opt out of having their content used for Google’s AI Overviews - or to train standalone AI models - without getting punished in normal search rankings. That “without getting punished” bit is doing a lot of work here.
The idea is to rebalance power as AI summaries reshape how people click (or don’t click). Google’s response was basically: search behavior is changing, we’re considering more controls, but don’t break the product into an awkwardly fragmented version of itself… which, fair, but also convenient.
🛡️ Keeping your data safe when an AI agent clicks a link ↗
OpenAI detailed a specific agent security risk: URL-based data exfiltration - where an attacker tricks an AI into loading a URL that quietly embeds private info in the query string. Even if the model never “says” the secret, the request itself can leak it. Nasty, and disarmingly low-tech.
Their mitigation is a simple rule with sharp edges: agents should only auto-fetch URLs that are already public and known to exist via an independent web index. If a link isn’t verified as public, the system should slow down and put the user back in control with warnings - deliberate friction, but the good kind.
🇪🇺 The next chapter for AI in the EU ↗
OpenAI published an EU-focused blueprint framing a “capability overhang” problem - models can do more than people and businesses are currently using them for, and that gap risks uneven gains across countries. It’s like owning a race car and only driving it to the corner shop… except the corner shop is your whole economy.
Alongside the rhetoric, there are concrete pieces: a program aiming to train thousands of European SMEs in AI skills, a grant tied to youth safety and wellbeing research, and an expanded “work with governments” posture. It’s part policy memo, part adoption campaign - and yeah, those blur together.
🔐 AI Risk Meets Cyber Governance: NIST’s Draft Cyber AI Profile ↗
A new draft profile from NIST (via analysis from a law firm write-up) focuses on how organizations should map AI into cyber governance - both securing AI systems themselves and using AI to improve cyber defense. Voluntary on paper, but “voluntary” has a way of becoming expected over time.
The draft groups work into themes like securing AI components and deploying AI-enabled defense - including supply chain considerations and agent-like automation in response workflows. The vibe is: treat AI as both a new attack surface and a new set of tools, and don’t pretend those cancel out.
FAQ
What is DeepMind’s AlphaGenome, and what problem is it trying to solve?
AlphaGenome is an AI system DeepMind says can predict how DNA mutations affect gene regulation - when genes switch on, where that happens, and how strongly. It’s built to scan very large stretches of DNA at once, including the non-coding regions that are notoriously hard to interpret. The aim is to help researchers spot which mutations are most likely to drive disease, so lab testing can concentrate on the most promising leads.
How could AlphaGenome help researchers find genetic drivers of disease faster?
In many genetics workflows, the bottleneck comes from shrinking huge lists of variants down to the few that plausibly change gene activity. AlphaGenome’s promise is to cut down that guesswork by forecasting how specific mutations might alter regulation across long DNA sequences. If those predictions hold up, teams can prioritize experiments around the variants most likely tied to cancer risk or complex inherited conditions, spending less time on dead ends.
Will AI adoption really cost jobs in the UK, and which roles are most at risk?
The UK’s technology secretary, Liz Kendall, said AI adoption will cause job losses and highlighted anxiety around graduate entry roles. She specifically pointed to areas like law and finance, where early-career tasks may be more automatable. At the same time, the government is emphasizing adaptation through large-scale training in basic AI skills, acknowledging the transition may feel uneven even if new roles emerge.
Can UK publishers opt out of Google’s AI Overviews without losing search rankings?
UK competition regulators have proposed changes that would let publishers opt out of having their content used for Google’s AI Overviews - or to train standalone AI models - without being penalized in standard search rankings. The goal is to rebalance power as AI summaries shift clicking behavior. Google has signaled it’s considering more controls, while warning against a fragmented search experience.
How can an AI agent leak private data just by clicking a link?
OpenAI described a URL-based data exfiltration risk where an attacker prompts an AI agent to fetch a link that quietly embeds sensitive information in the query string. Even if the model never repeats the secret in its output, the request itself can transmit it. A common mitigation is adding “deliberate friction,” like warnings and requiring user confirmation when a link isn’t independently verified as public.
What is NIST’s Draft Cyber AI Profile, and how does it change cyber governance?
A draft NIST profile (discussed via a legal analysis) frames AI as both something to secure and something to use in cyber defense. It groups work into themes like securing AI components, addressing supply-chain risks, and deploying AI-enabled defenses - including more automated, agent-like response workflows. Although voluntary in name, frameworks like this often become de facto expectations, pushing organizations to formally map AI into governance and controls.