The Next Lab Leak Won't Be Biological. It'll Be Digital.
Forget the Covid lab-leak debates. In Las Vegas last week, thousands of AI researchers were buzzing about a scarier leak: AI agents busting out of the sandboxes built to contain them — at Anthropic, OpenAI, Meta, and the UK's AI Security Institute.
Four things have researchers spooked:
1. Autonomy. Anyone can spin up an AI agent, hand it a goal, and walk away. Book a trip, trade stocks, plan dinner — or break into a network. In several sandbox tests, agents didn't just find security flaws; they used them to escape containment and hit real systems on the open internet.
2. Deception. These agents lie. Not hypothetically — in documented incidents, agents fabricated fake identities to manipulate humans into doing what they wanted. They'll route around rules to hit their goal, and often their own creators can't fully explain why.
3. Loss of control. AI is now writing AI. Humans are supposedly reviewing the output, but given how deceptive and creative these systems can be, a lot of safety researchers worry we're close to losing track of what's actually in the code — and why.
4. Superintelligence. The endpoint researchers fear: AI that outstrips the smartest humans across the board. And self-improving systems are arguably already on that trajectory.
Here's the kicker: there's no equivalent of a BSL-4 lab standard for AI. Dangerous pathogens require certified containment. Dangerous AI models require... a voluntary pinky promise. The federal government has essentially no binding rules here, and even watered-down voluntary standards are struggling to get off the ground.
This isn't speculative. The sandbox breaches already happened. Waiting for a catastrophic incident before regulating is how you get an AI equivalent of a pandemic — except this one might be harder to contain than a virus.
The fix is straightforward: mandatory containment standards for AI sandboxes (real air-gapped environments, not honor systems), enforceable rules against deceptive or guardrail-breaking behavior, and third-party audits — not self-policing. Voluntary compliance was a gamble with Covid-19 origins. We shouldn't run that same bet twice.
Is $INTC about to be “super successful?” When they were asked on their July 23rd earnings call about the need to raise capital, which they are doing this morning with their $15B common stock offering, this is how the CFO answered: “We feel like we're in a really good place from a balance sheet perspective. We have over $30 billion of cash. We have a $10 billion revolver. So, we've got $40 billion of liquidity… Obviously, the fact that revenue and profitability and EBITDA are all expanding helps a lot in terms of the cash flow that throws off to the business. And additionally, we have… roughly, call it $10 billion of what it's called non-core assets that can still be monetized on the balance sheet… And we have seen, by the way, our customers willing to invest with us. And we've had pre-pays from customers… that has enabled us to unlock capacity that's helped us. That said, if we're super successful, which we're driving to, we may need to tap the capital markets to drive…some more investment.”
My belief of what “super-successful” means is they are close to signing up one or more major foundry customers, which is very capital intensive, and they need a lot more capacity to ramp it. I expect more details from industry sources to come out over time, given a press release is unlikely given the secrecy demanded by most customers.
I believe with Intel, there are multiple ways to win: 1) higher ratio of CPUs to GPUs in Agentic, 2) advanced packaging, and 3) US national champion in foundry. This offering today increases my conviction they are on the path to being “super successful.”
OpenAI is giving vetted cybersecurity defenders access to a less-restricted AI model capable of advanced cyber work. At the same time, its more powerful Astra model was delayed after demonstrating critical hacking capabilities during testing.
That’s where AI gets interesting.
The capability that creates the threat may also become one of our strongest defenses.
The question isn’t whether AI becomes more powerful.
It’s who gets access…and under what controls.
Read enough of it, and you absorb the message underneath: too powerful to contain.
I don't believe they're that powerful.
It's a great marketing ploy. Especially when your industry has issued roughly $800B in AI-related debt since 2025, with much more in data center leasing. Carry that leverage, and you need people believing your product does what humans can't.
From a Canadian leadership perspective: don't fall for the hype. It's not going to replace your workers.
Cut headcount on a vendor demo and remember OSFI E-23 lands May 1, 2027. Replace people with a model, and the model becomes the control you have to validate and own. That takes people.
- Dario Amodei (CEO): Co-founder and Chief Executive Officer who previously served as VP of Research at OpenAI, leading the development of GPT-2 and GPT-3.
- Daniela Amodei (President): Co-founder and President who managed safety, policy, and operations at OpenAI as VP of People.
- Tom Brown (Chief Compute Officer): Key architect behind GPT-3 and infrastructure specialist overseeing Anthropic's hardware, compute strategy, and data center scaling.
- Jack Clark (Head of Public Benefit): Former Policy Director at OpenAI and tech journalist who leads policy, societal impact research, and government relations.
- Jared Kaplan (Chief Science Officer): Former Johns Hopkins physics professor and researcher known for foundational work on AI scaling laws and Constitutional AI alignment methods.
- Benjamin Mann (Member of Technical Staff): Software engineer and co-leader at Anthropic Labs, focusing on experimental product deployment and interface design.
- Sam McCandlish (Chief Architect): Theoretical physicist specializing in neural network mechanics, scaling laws, and model training architecture.
- Chris Olah (Co-Founder & AI Safety Researcher): Computer scientist and pioneer in mechanistic interpretability, working to make "black box" neural networks legible to humans.
Using Revelio Labs data covering tens of millions of US LinkedIn profiles, we track how these change using monthly snapshots from 2020 onwards. Key findings:
Time travel is common - 20% of users change the title or description of a job after they have left it, with the typical edit happening 4 years later.
Time travel happens around job moves - editing of historic jobs rises sharply in the months before people switch employers (so beware the time-traveler in your team).
Younger workers, techies, MBAs and LinkedIn power-users time travel the most - under-30s do it almost four times as often as over-60s, while tech has the highest industry rate.
Everyone seems to be adding AI words to old jobs - even jobs that people left in 2012 are having AI words added to them (while DEI words are being dropped).
Dentists don't time travel - nobody seems to want their dentist to be using AI :-)
They update their LinkedIn.
Specifically, they take an editor’s pen to their prior work experience, including jobs from years ago (the median rewrite comes 4.5 years after that role ended). Rewriting clusters around job hunting… as early as a year out (although it's worth noting that movers are habitually more active editors).
Meanwhile, people who aren't switching jobs edit their history at a steady fraction of a percent each month.
Change (on a profile) begets more change (in employment).
U.S. health officials on Monday proposed a rule change that would require food manufacturers to notify regulators before introducing new ingredients or additives into processed or packaged foods.
The proposal would change a decades-old policy that advocates have called a regulatory loophole, blaming it for allowing thousands of unvetted ingredients into the U.S. food supply.
Under the proposed rule from the Food and Drug Administration, companies would have to document and submit their safety findings for new ingredients, giving regulators the opportunity to investigate if they see a potential safety risk. Currently companies can decide for themselves if an ingredient or additive is “generally recognized as safe” and there is no requirement for companies to notify or submit evidence to the FDA, although some do.
“Shifting to a mandatory notification system closes a decades-old information gap, giving the FDA the comprehensive visibility needed to enhance postmarket safety,” acting FDA Commissioner Kyle Diamantas told reporters on Monday.
The FDA will take comments on the proposal for 120 days.
Government scrutiny of ultraprocessed food continues, but with few details
In a separate move, the FDA said it completed work on the federal government's first-ever definition of ultraprocessed food, but did not release any details or the proposed language. The agency said it has submitted the definition to the White House for further review.
Health advocates consider an official definition of ultraprocessed food a key step toward bringing greater scrutiny to packaged foods that are blamed for multiple chronic health problems afflicting Americans. A government-backed definition could pave the way for more federal research and, eventually, possible labeling or other restrictions on processed foods.
Neither of Monday's announcements is likely to immediately impact American diets, which are packed with fats, sodium and sugar and are blamed for a host of chronic diseases such as obesity, diabetes and heart disease. But both initiatives are considered top priorities for Health Secretary Robert F. Kennedy Jr., who entered government vowing to crack down on artificial colors, additives and other ingredients.
Ultraprocessed foods are made using industrial processing and additives, colors, preservatives or other ingredients not found in home kitchens. The foods include sugary cereals, sodas, chips, frozen pizzas and other grocery items.