
It has been a rough few weeks for computer systems. None of this week’s outages has been tied to an attack. That is almost the point: the systems fell over on their own, and separately, the attackers stopped needing us.
Information systems are best understood by the term CIA. The CIA triad is the three pillars of information security: confidentiality, integrity, and availability. Secrets exposed, data you cannot trust, or a system that is offline are different examples of a breach. These outages are a story for another time.
On Monday, Google’s Threat Intelligence Group published something it had not seen before. Russian-speaking operators ran a cyberattack with no human in the loop at any stage. A system of chained AI agents found the hole, wrote the exploit, and fired it, start to finish, on its own. The targets were French media and education sites.
The same Monday, California’s attorney general sued OpenAI over a summer in which its agents went hacking. The question for anyone funding this is no longer whether machines can attack alone. It is what defense is worth once they can.
The Deep Dive
Patches are now blueprints
The centerpiece of the Google report is a tool its researchers called exploit_agent. It chains several AI agents with frontier and locally hosted language models. It watches widely used software and compares the code before and after a vendor ships a patch. From the change it infers what the patch is hiding, writes an exploit, and runs it. No analyst reads the diff. The machine does.
That inverts the oldest rhythm in security. A patch used to buy the defender time. You shipped the fix, and attackers had to study it, understand it, and build a weapon before enough targets updated. The gap between patch and exploit was the defender’s margin. Google’s finding is that an agent now closes that gap in hours. After two PHP vulnerabilities were patched on October 21, hundreds of exploits appeared within days, most of them nonfunctional and most of them, Google believes, written by AI. Shipping the fix shows the attacker where the hole was.
Why speed decides it
Google counted 22 exploited zero-day flaws in August alone. That is up from 11 a month earlier in 2026 and an average of 8 a month in 2025. The researchers found the same automation in open-source offensive frameworks, including two named CAIRN and PentAGI, used against a US aerospace company and organizations across Asia. A China-linked group is building its own automated framework for espionage.
The pattern is in the numbers. When the cost of an attack falls toward zero and the time to build one falls toward minutes, volume stops being the constraint. The attacker who can run the fastest loop wins, because the attacker who finds the weakness first exploits it before anyone patches. This is not a new dynamic in markets. It is what happened to trading.
What high-frequency conflict looks like
When US equity markets automated execution after Regulation NMS in 2005, the advantage left the smart analyst and went to the fast machine. Firms paid for co-location and fought over microseconds. On May 6, 2010, the loop ran away from everyone and wiped out close to a trillion dollars of value in minutes before it bounced back. The exchanges did not respond by hiring more humans. They built automated circuit breakers that acted at machine speed, because nothing slower could keep up.
Cyber is entering the same regime. The contest is loop speed. A defense that waits for a human to read an alert, decide, and respond is a defense running at human pace against an attacker running at machine pace. It loses the race it is in. Corma came out of stealth in August. This week it raised $125 million from NEA at a $1 billion valuation, on a pitch that is a defensive foundation model acting on its own. The thesis is autonomy on defense, because autonomy on offense is now shipping.
The Investment Takeaway
New North Ventures believes the defensible cyber businesses are the autonomous ones. A tool that flags a threat and waits is becoming a liability the moment the threat moves at machine speed. The value compounds in detection and response that acts without a person in the path, not in another perimeter product. That is where we underwrite, and we weigh a company by the speed of its loop before its headcount. Two of our portfolio companies sit on this line. Reality Defender scores synthetic voices and faces in real time, inside the call rather than after it. CrunchAtlas runs an active-defense loop on critical infrastructure, built to act on its own. This is not a forecast. We are funding the companies that answer it at the same speed.
The Funding Ledger
UFORCE raised $700 million at a $6 billion valuation (Series A), led by the UK National Wealth Fund and Valor Equity Partners. London maker of unmanned air, sea, and ground systems and battle-management software.
Onodrim Industries raised $300 million at a $1.3 billion valuation (Series A), led by Lux Capital. Amsterdam platform buying aerospace, defense, and industrial manufacturers to build a European defense base.
Corma raised $125 million at a $1 billion valuation (Series A), led by NEA. Tel Aviv lab training a defensive cybersecurity foundation model that runs automated agents against multi-stage intrusions.
SE3 raised $30 million (Series A), led by Accel. Munich maker of spatial-AI software that lets one operator command a drone swarm by voice, with no GPS.
Primordial Labs raised $18 million (Series A), led by Ballistic Ventures. New Haven maker of software that lets an operator task a drone by voice, like a teammate.
Three of the five rounds are European, and the two largest, UFORCE in London and Onodrim in Amsterdam, took a billion dollars between them. The lone American round, Primordial, is the smallest on the list. Four of the five fund uncrewed systems or the software that flies them. Scaled dual-use capital is moving to European sovereignty plays while US venture writes the small early checks. Nobody wrote the big American hardware check this week.
The Northeast Desk
The region’s freshest venture news is in the ledger above. Primordial Labs, in New Haven, raised $18 million to let a soldier run a drone by talking to it. Google documents attackers pulling the human out of the loop. A Connecticut company just raised money to keep the human in it, in command of the machine by voice. CEO Lee Ritholtz has moved the product, Anura, from drone demos toward armored-crew interfaces.
The region’s deeper asset is the workforce that builds the fleet. In August, General Dynamics Electric Boat committed $25 million to a new training center in Warwick, Rhode Island. The Navy’s Submarine Industrial Base program matched it with $20 million. The partner is the New England Institute of Technology. The school expects about 3,000 workers to move through it over three years, then head to Electric Boat’s yard at Quonset Point in North Kingstown. The submarine base is the one piece of defense manufacturing the Northeast cannot lose, and the bottleneck has always been people, not orders.
Two events worth putting on the calendar: MIT Lincoln Laboratory, in Lexington, Massachusetts, holds its KASPR Technology Interchange Meeting on December 11. The same lab runs its Artificial Intelligence for Cyber Security Workshop during the AAAI conference, February 16 to 23, 2027. Lincoln Laboratory is where New England’s defense research and the deep dive’s subject meet, and both sessions are open on its calendar now.
Portfolio Corner: Hubble Network
In late September, Hubble Network closed a $200 million Series C at a $1.6 billion valuation, led by Smith Point Capital, which made the Seattle company a unicorn. We featured Hubble in July, before the round, so treat this as the update: the round is done, and it was big.
Hubble builds a satellite network that any Bluetooth device can reach directly, with no cell tower and no new hardware on the device. A standard Bluetooth chip, the kind already in billions of trackers, sensors, and tags, talks to a satellite overhead. The company spent the year widening that aperture, from a Texas Instruments collaboration in January to opening coverage to all Bluetooth devices alongside the raise. CEO Alex Haro, who co-founded Life360, is running it.
The dual-use angle is simple. It is connectivity where there is no grid. An asset tag phones home from open ocean. A sensor reports from a remote perimeter. A device holds a location fix where GPS is jammed. All of it runs on hardware that already exists and a signal that comes from orbit. The defense logistics problem is rarely the tracker. It is coverage. Hubble is selling coverage as a layer you do not have to install. New North Ventures is an investor. A $1.6 billion valuation says the market believes it.
More links to explore
Google’s Threat Intelligence Group on the first fully automated AI attack, via CyberScoop.
The Register‘s technical read on the agent that turned a patch into a working exploit.
California sues OpenAI over its agents’ summer hacking spree, via The Record.
Hubble Network’s $200 million Series C, via Payload.

Wouldn’t it be more accurate to say the outages have been acknowledged as being perpetrated by a foreign actor?