@ConorBronsdon@CornelissenJo@DataCamp Writing the next answer while the learner is still thinking is a clever trick. It's like a waiter who starts your usual order the moment you walk in, so the wait just disappears. I'm curious how often that guess turns out wrong and the work gets thrown away.
@iamchrisani I agree the tool doesn't decide the quality, the checks around it do. When I was teaching an AI to write Rust, what made me trust it was an exam of hidden tests it never saw. It passed 59 of 59, and without that I couldn't tell good code from code that only looks good.
That matches what I'd guess. Binding the input, meaning the user's text goes in as data and never gets pasted into the query, is a pattern these models have seen millions of times. Who is allowed to call the endpoint is different, because that rule lives in your head, not in the code. Did either agent add a login check without being asked?
Lovable is an app that builds a whole website from a plain-English description. Many of those sites keep their data in Supabase, a hosted database that the web page talks to directly.
A German security firm, Crissy Field, looked at 159,041 Lovable projects. On 24,664 of them, the page pulled database records before anyone logged in. That part can be fine, since some data is meant to be public.
The problem came next. When they asked the same database for a bit more, 7,403 of those sites, about 30%, handed back records with email addresses. One service exposed more than 1,000 ID records, passport numbers included. The researchers are careful to say that 30% is not a confirmed leak rate, since some of those emails are just public contact addresses.
Lovable itself wasn't hacked. Each app's own access rules simply let strangers read too much. The fix is checking who is allowed to read each table before real customer data goes in.
It's like getting a beautiful new house where the locks are fitted but nobody ever turned the key.
I'm building a self-hosted Rust CMS, and this is the check I'd put first open every page logged out and see what comes back.
Have you ever tried using your own app while logged out?
@almokhtar_al It's like letting someone write their own permission slip. The rule I keep coming back to is that the file setting an agent's limits should be read-only to the agent, and any change should need a human. If the agent can edit its own rules, they're really just suggestions.
@pauliusztin_ I like that the cap lives outside the agent, where the model can't talk its way around it. A limit written in a prompt is a request, while a limit written in plain code is a wall. Do you track how often the cap actually fires?
The banks are calling it a technical problem, not an attack. When several big banks break in the same hour, my first guess is something they all lean on, like a shared login or payment provider, rather than four separate bugs. That's the scary part of modern systems you can test your own code all day and still go down because of a piece you don't own.
A dormant defect is a bug that sits quietly until the right conditions wake it up. What gets me is the failover part. That's the system meant to move traffic away from trouble, and here it helped spread the problem while there was less spare room. Safety systems need testing too, like checking that the fire exit actually opens.
I'd put step 4 first. Evals are just the tests that tell you whether the agent actually did the job, and without them the other three steps are guesswork. When I tested an AI on Rust, I kept the tests hidden from it, because a model that can see the test can learn to pass the test instead of solving the problem.
Antithesis runs your whole system in a simulated world where it can replay the exact same failure as many times as you want. For me, that replay is the big deal. A crash you can't make happen twice is the worst kind to fix, because you can never prove the fix worked. How does the agent decide which failures to try first?
My brain has a list of excuses ready every time I try to start something new.
"Not ready." "Tomorrow." "Someone already built it."
I wrote about why it does that, and what finally gets me moving anyway. https://t.co/I8b8WPTXd4
Giving the agent its own identity also fixes the audit problem, because you can finally tell its actions apart from yours. A key that expires in 15 minutes is also far less useful to whoever finds it later in an old log. Did you also have to clean secrets out of the agent's own transcripts?
The scary part is that the monitoring tool is the thing lying. nvidia-smi, the little program that shows how busy your GPU is, reads from a library the malware quietly replaced. That's why I think a monitor should compare two signals that are hard to fake together, like reported usage and power draw. If a card says 0% but is pulling full power, something is off.
I think that's the part people skip because prefill numbers look so good on a chart. Decode is where the model writes one word at a time and keeps going back to memory for each one, so memory speed decides how fast it feels. Until someone shows decode results across a few different models, I'd keep CUDA in the toolbox.
Google's Project Zero is a team that finds security bugs in other companies' software. For years they gave a company 90 days to fix a bug, then 30 more days for people to install the fix, before sharing the details.
Last week they dropped those extra 30 days. Their reason AI now helps attackers look at a fix and work out the bug behind it almost right away. The 90 days also drops to 60 in February, then 30 in 2028.
It's like a locksmith recalling a lock. The day the recall goes out, burglars know which doors to try.
I'm building a self-hosted Rust CMS, so soon this one is on me. How long do updates sit before you install them?
The simple version before training starts, every weight gets a random starting value, and how those random numbers get laid out depends on how many cores the chip has. So the starting point quietly records the machine, like a potter's thumbprint left in the clay. The careful part matters though. It tells you which chip started the run, not where the chips sat or who paid for them.