A lot of teams underestimate how different backup and restore are.
Backups can run perfectly for months, but that doesn’t mean you can recover in hours.
I’d encourage every engineering team to actually test a full restore with production-sized data. You might be surprised.
The attacks this week have one thing in common.
AI has made it cheap to run the same attack against hundreds of companies at once, so more of them are going to get through, including at companies that patch and detect well.
Prevention is critical, but no company can count on stopping every attack.
You also have to know that when one gets through, you can get your data back clean and be running again in hours.
If you haven’t yet:
• Check that every cloud resource is actually backed up. Most teams find the gaps in the middle of an incident.
• Keep backups in a separate account that stolen credentials can’t reach or delete.
• Know your last clean point in time before you need it.
• Run a real restore and time it.
I think a lot of companies are going to find out the hard way that detecting an attack and being able to recover from it are two very different things.
Agents part is underrated here. Humans found a fax machine, an agent that loses its mailbox just stops and imo most teams still haven’t tested what theirs does when that happens
Last week’s Microsoft 365 outage was the first of many, and the next ones will be worse.
The way companies depend on their data has changed, and the way most of them protect it was built for a different time. https://t.co/BiGMmo2wmY
Congrats @pluraiAI ! This is really important, agents behave in surprising ways and ignore instructions in enterprises https://t.co/PyB6Cf4Xng - looking forward to our collab around https://t.co/Ph7L8ou9IN
Excited that we're research partners @Eon_io_ is and congrats on this!
Plurai Simulation is the #1 Product of the Day on Product Hunt 🏆
Thousands of builders showed up, not just to test their agents, but to break them on purpose, before their users do.
Because agents don't fail in the demo. They fail on the 10,000th conversation, in the case nobody thought to test.
That's why we built Plurai Simulation the Waymo way: simulate real-world scenarios at scale, find where your agent breaks, and fix it before production.
Until recently, this was available only to a handful of enterprise teams. Now every engineer can run it on their own agent, right from their coding agent.
Thank you to everyone who upvoted, tried it, shared it and pushed us with feedback. This is just the beginning!
Try it for free: https://t.co/Nqire8qS3j
@LucidCap@urieli17 Great question! Co-Author here - It's literally impossible for us given the permissions we in take in AWS. Customer data is sacred in Eon we designed our system so that we have no access to it.
Era by Eon uses usage metadata on specific select opt-in customers.
@AssafPetronio@OfirEhrlich In a tl;dr; sketching algorithms to collect real joint distributions, real usage metadata and llm adversarial GAN like runs at several stages
@nahid_pro09@OfirEhrlich Yes! That's a fundamental and (really cool) part of the design. Promise to write a blog post about this you can check the published research (in arxiv) about this in the meantime https://t.co/yGTQkxliRp !
@catmanyau@OfirEhrlich Hey, worked on this - Nothing beats production data! We live in a world customer data is strictly off limits and inaccessible to engineers. We never train on or look at customer data and architecture wise it's impossible for us - Era closes a real gap we've had for years!
@DakshTrehan@OfirEhrlich That's the cool part! We post-train on one company (well, tens) and validate on another and we prove statistically that we're not overfitting on the company we trained on! How did you guess we do post-training and alignment with this that's a good guess you caught us!
@oleginprod @OfirEhrlich Yes! There is a churn specific data overlay and there is an "outdated board" overlay. We combine them internally for some exciting soon-to-be-released stuff - hit us up if you want early access!
@DocumentingAGI@OfirEhrlich Hey! This is really interesting - basically it stays very similar though not an average. Eon never looks at customer data (obviously) so this is inferred strictly from usage metadata and anonymized sketching.
@paulhq@OfirEhrlich Lots of things, as a simple example agents will find an old JIRA board and assume it contains up to date data, find an excel file describing revenue and assume it contains the correct up to date information or find identies and mix them up
@omarsar0@OfirEhrlich Hey, I worked on this - you can! This is called an "overlay" in Era speak.
There are several ways to create "messy" data in Era and we generate data - the business scenario picker is the simplest.
We also have really interesting research papers about split identities coming!
@johnroodepic@OfirEhrlich Hey, worked on this - yes it can. You can inject latency and faults and it's based on real faults from production. That said - Era online (SaaS) is readonly so we can withstand the scale of offering this for free :)
@oriSomething מניסיוני לא רק שמאוד קשה לשכנע אנשים ללמוד את המקצוע שלהם טוב, זה הולך ונהיה יותר גרוע עם הזמן. אתה רואה אנשים שלא מכירים כלים לחלק ביצועים שמתייחסים לזה כמו מיתוס ומנחשים כאילו הכלים לבדוק את זה לא סופר זמינים והם לא יכולים פשוט להעלות לוג לsystem analyzer ולראות למה זה איטי
@dan_abramov Also Jake's comment (above) is very authoritative IMO, he's generally very trustworthy/knowledgable in all things related to AbortSignal/fetch.
@dan_abramov Was pinged here - In normal conditions yes, naturally it's not "ironclad", someone can always override window.fetch or `AbortController` or Promise.prototype.then or other stuff. Feel free to open an issue in whatwg/fetch and I'll happily list more specific guarantees.
That said, if you have a strong opinion about what Node.js should do regarding cancellation please _do_ email me (email can be found in the Node.js repo readme) or send me a facebook message. I am happy to chat I just don't like the violence of twitter.