📘 Unsupervised Studying NO. 507 – Cyber Tech

My buddy at OpenAI is hiring offensive safety engineers to pink crew their AI brokers My buddy is an absolute star, and I’ve heard nothing however good issues in regards to the crew. If that is you, you must examine this out. OPENAI JOB POSTING

Tremendous excited to be consolidating and upgrading numerous my enterprise workflows earlier than heading into 2026. I’ve a brand new Head of Operations who’s extraordinary, and we’re consolidating on a single instrument (reply to this electronic mail and ask for the title in case you’re curious) for each mission administration and chat.

I additionally discover it humorous that the instrument simply applied Slack as a minor function throughout the platform. Like I actually don’t have any function inside this platform’s chat that I haven’t got in Slack. However the reverse is certainly true. Truthfully glad to see it as a result of I really feel like Slack has handled folks badly simply because they might.

Huge improve to my voice system for Kai. Now I am utilizing the upgraded API endpoint for Eleven Labs. So all of my completely different brokers, together with Kai and all of the sub-agents, are utilizing parsody markers for his or her voices, which is mainly a markdown system for indicating emotion.

I am undecided what’s occurred within the final couple of months, however it seems that X has largely made a return as the primary place for dialog on-line—no less than inside tech, safety, and positively inside AI. I’m now recommending to mates that left—lots of which for good causes—take into account coming again. If you happen to actually have a look at them, any platform that you just take part in goes to have folks and concepts related to it that you just deeply dislike. And sooner or later, it is not a internet win trade-off for you, or anybody else, to exclude your self from dialog and exercise due to that. I personally have seen the negativity and hate stuff drop off considerably in the previous couple of months, and there are indicators that that’s going to enhance. This mixed with my liberal use of filters and muting, and the truth that the choice have largely self-immolated, have introduced me to this conclusion. Simply take into account it.

I upgraded my dictation microphone to the up to date model of my Sennheiser growth mic, the MHK 8060, significantly to be used with Wispr Circulate. No, they aren’t a sponsor, but when they’re listening, they need to hit me up to allow them to be. 🙂

I used to be beforehand utilizing a high-end devoted Shure mic for this, however these are Condenser mics, that are designed to be actually clear, however typically aren’t as delicate. The Sennheiser growth mics are extraordinarily delicate, and the entire sport with dictation is I simply wish to kind of discuss and converse naturally as if I am speaking to a colleague, and since I am doing large dictation numbers I do not wish to should yell or make numerous corrections. The 8060 is overkill as a result of it is a skilled mic, however in case you get any dynamic mic, together with like a Yeti, it should be approach higher than dynamic mics for this specific use case. The factor with a Yeti is you would want it to be on an arm or proper in entrance of you in some way, near your mouth so you’ll be able to discuss quietly. That is why I just like the growth possibility so it’s sort of out of the way in which usually.

Anyway, in case you’re doing numerous AI, particularly utilizing dictation, I extremely advocate you turn to a dynamic microphone. Ideally, a growth mic that’s pointed in direction of you for doing AI dictation. And in case you’re doing numerous AI whereas typing, I like to recommend you turn to dictation. 🙂

Huge updates to the Private AI Infrastructure Mission.

🔥 Greatest story for me final week was the the Anthropic report on Automated Cybersecurity Assaults. The quick model of that is that Anthropic discovered one other automated hacking marketing campaign utilizing their platform, and so they wrote a full report on it.

We have been saying right here for a very long time that the way forward for cybersecurity is attacker automation methods in opposition to offender automation methods. And now we’re seeing that truly come true

To me, this additionally scores as a win for scaffolding over mannequin as a result of evidently the primary approach that it labored was breaking issues into these discrete items as a part of an orchestration system

The one cause that we have now any visibility into this factor in any respect is as a result of they are not utilizing customized fashions. What occurs when China itself has fashions that attackers can use which can be pretty much as good or higher than this present technology?

Maybe the scariest factor to me is that the open-source fashions or the proprietary fashions utilized by China solely should get a sure degree of high quality earlier than the scaffolding itself—which will be shared as a sequence of expertise or as a whole automation system on GitHub or one thing—will be distributed to all attackers inside their ecosystem and utilized in assaults at scale. (AI won’t power me to cease utilizing em-dashes)

For the previous couple of months, I’ve really began to be involved about cybersecurity for the primary time shortly. Like I’ve at all times identified that the assaults will proceed to evolve, they will proceed to outsmart the defender, and the defender will then evolve. So it is going to proceed to be a Crimson Queen sport.

I am beginning to get actually involved that defenders can be too far behind on this maturity mannequin that I simply put collectively for my current discuss at AppSec USA.

This hole between defenders and attackers on this maturity mannequin is absolutely regarding to me. Primarily based on what I am seeing and listening to from the sector, I’m fearful attackers are a lot additional alongside at implementing Stage 2, and a few are already transferring in direction of Stage 3, whereas most defenders are nonetheless at degree one and simply beginning to consider degree 2.

And I am not simply speaking about start-ups or common firms or giant enterprises. I am additionally speaking in regards to the US, versus China, Russia, North Korea, or whoever. And to some extent even of a lot smaller actors who’re simply actually superior orchestration at degree 2 – 4.

My concern is that I have been in too many a whole lot of firms at this level serving to them with safety. And I’ve by no means seen anybody who can radically shift their defensive posture to account for a brand new risk this shortly.

Whether or not you are a enterprise competing in opposition to rivals or an organization defending in opposition to attackers, the one approach to beat a system like that is to have a system like this.

And your system must be higher. It wants a greater design, must be sooner, wants higher orchestration, wants extra context, wants extra real-time knowledge, and many others. It is just like the cat-and-mouse sport that we have all been speaking about for many years is now much more concrete, and it is personified as an AI orchestration system.

Attackers have the large benefit of having the ability to YOLO this sort of factor. They’ll launch 10 of them and see which one works. And if it doesn’t work and sort of breaks all the pieces, they simply delete and begin over. They do not have as a lot SRE to fret about and 5 nines to keep up and 37 completely different approvals earlier than they’ll even begin this system.

Defenders actually are in a foul spot right here, and that is the one tech struggle we can’t lose.

Logitech obtained hit by a zero-day exploit that leaked 1.8 terabytes of knowledge The Clop ransomware gang used an Oracle vulnerability to brush inner knowledge, although Logitech says no delicate buyer information was compromised. SEC FILING | LOGITECH DISCLOSURE | BLEEPINGCOMPUTER REPORT

Attackers now weaponize half of all CVEs inside 48 hours utilizing AI Gene Moody at Action1 says defenders cannot win at human pace—automation is not non-compulsory anymore, it is the one approach to shut the hole. Utterly agree. That is what we have been saying right here for a few years. Much like what we talked about within the first story. ACTION1 HOMEPAGE | ACTION1 PATCH MANAGEMENT | CISA KEV CATALOG

China threatens Japan with navy patrols and vacationer warnings after Taiwan protection feedback After Japan’s new PM Sanae Takaichi stated she’d defend Taiwan militarily, Beijing despatched warships to disputed waters and informed Chinese language vacationers to keep away from Japan completely. OODALOOP ARTICLE

Poland says two Ukrainians working for Russian intelligence sabotaged railway then fled to Belarus Polish PM Donald Tusk informed parliament that investigators recognized the suspects behind the rail explosion he referred to as unprecedented sabotage. POLISH RAILWAY EXPLOSION REPORT | ORIGINAL SABOTAGE STORY

America and China are neck and neck in navy drone capabilities A head-to-head comparability reveals each superpowers have roughly equal unmanned plane arsenals. WSJ DRONE COMPARISON

Europe plans a drone wall in opposition to Russian incursions however consultants say it will not be foolproof Safety correspondent Frank Gardner stories that after Russian drones flew into Poland, Europe desires a multi-layered protection system by 2027—however the tech won’t ever be watertight. BBC INDEPTH ARTICLE

Google launched Gemini 3, and the metrics point out that it is a main leap However the actual check will come from folks really utilizing it within the area. I’ll be incorporating it in my Kai system, after all, however I do it via expertise and particular calls to the Gemini command line. a few of these jumps over earlier fashions are extraordinary, particularly Humanity’s final examination.

We obtained GPT-5-1 at 26% and 3Pro at 45%. From 17% in Arc AGI-2 to 31%. Have a look at Math Area Apex. That one is ridiculous. GPT-5-1: 1% to 23%. I assumed this merchandising bench 2 metric was actually cool. That is mainly an precise merchandising machine enterprise that is being run autonomously by AI. GPT-51 made $1,400 in revenue, and Gemini 3 Professional made $4,600 in revenue.

details Benchmark Suite from 50% to 70%. Easy QA verified 34% to 72%. Actually spectacular numbers. I simply cannot wait to see how this performs out in real-world situations.