Summary
Audio Summmary
Much focus remains on AI-enabled cybersecurity attacks. In traditional human-led cyberattacks, the criminal launches perhaps dozens of commands in an hour. An AI agent orchestrating a cyberattack can launch thousands of commands. For instance, the FBI is warning that Artex, an open-source AI hacking tool, is being used since July to launch autonomous-driven cyberattacks around the world. Targets include financial services, government, Tech companies, defense, phone and internet companies, and law firms. The Palo Alto Networks’ Unit 42 organization writes that in some cases, “attackers needed just 72 minutes to move from initial access to data exfiltration, 4 times faster than last year”.
Speaking at the recent New York Times’ DealBook Summit, Microsoft CEO Satya Nadella said a lot of IT infrastructure in the US is running software that cannot be patched. As the speed of vulnerability discovery has increased due to AI, so has the speed of patching systems. This leads to precipitation in patching, where there is not enough time to thoroughly test the patch. This creates problems when the patch is made. Meanwhile, an InfoWorld article looks at governance principles for AI coding agents. As coding autonomy moves from the human to the AI, policy must no longer be simply a corporate document. Rather, policy must translate to engineering techniques that oversee agents. One technique is the creation of an audit trail for agents. Another is to maintain human accountability. For instance, agents may be allowed to generate code, but any change to the database might need human approval.
There has been much discussion following the resignation of Nvidia’s CEO Jensen Huang. TechCrunch examines the impact on AI startup funding. According to PitchBook, Nvidia participated in 105 rounds already in 2026. For example, Nvidia acquired Hugging Face for 12.93 billion USD in September. Huang’s investment approach saw companies requiring between 100 million USD and 500 million USD in GPU chips raise half of that amount from Nvidia – in return for use of Nvidia chips. This has helped to solidify the Nvidia ecosystem. Less startup investment from Nvidia could open the way for other companies to sell their chips. Google for instance announced two new chips last week. The Guardian newspaper looks into what is a very old problem of industrial espionage in the Tech world. The FBI believes that a number of North Korean agents are active in the US with the intent of spreading malware in cryptocurrency development companies. The nonprofit Center for Strategic and International Studies has identified over 224 incidents of Chinese espionage in the US between 2000 and 2022, half of which aimed to steal technology secrets. Many of these cases use a “honeypot-style” approach where the suspected spy pursued personal or sexual relationships with the target.
Another InfoWorld article considers the impact of misinterpreting headlines around the scarcity of compute power in data centers – and the impact this is having on organizational compute planning. In essence, capacity is currently tight as there is a high demand for GPUs, CPUs, and power supply. However, capacity is not yet sold out. Part of the issue stems from the cloud providers themselves who are communicating the narrative that they have demand for computing capacity that will last for years. This reassures their investors, but at the same time, confuses organizations into believing that future compute capacity is uncertain. Meanwhile, the International Energy Agency (IEA) has said that data centers will consume twice as much electricity by 2030 compared to today. Around the world, the next four years will see data centers generate 70% more carbon emissions as the US and China in particular use fossil fuels to power their data centers.
In relation to Big Tech, Nvidia reportedly told a US Congressional Committee that it is possible to introduce geo-tracking mechanisms into the AI chips. The mechanism, however, only locates GPUs. An extension is being envisaged where a chip’s license could be remotely verified during the boot sequence. The boot could be interrupted if the license cannot be validated. Microsoft CEO Satya Nadella justified an offer made by Microsoft to 60’000 employees urging them to quit instead of being part of a huge layoff. He suggested that the reason was AI – not that AI was replacing their jobs, but because AI was transforming workflows so job descriptions are changing. For instance, he said “software engineering is very much alive and, in fact, is going to thrive, except the way the team is organized and the work that a software engineer does has now got a new operating point”. Finally, Anthropic touched 2 trillion USD in market shares on its debut on the Nasdaq. The stock opened at 110.50 USD, 33% higher than the 83 USD IPO price. The IPO has been seen as a referendum on AI. Anthropic has nearly 900 billion USD in computing commitments with partners that include Amazon and Google. Its rapid growth depends on it being able to secure vast amounts of future computing capacity – which is not yet guaranteed.
Table of Contents
2. Jensen Huang’s departure blasts big hole in AI startup funding, VCs and founders say
3. ‘Jensen Huang lost his mind’: how AI’s smartest man fell for the world’s oldest trick
4. How AI agents are changing the rules of software governance
5. Panic over cloud capacity is out of hand
6. FBI says Artex AI open source hacking tool used to breach networks in ‘under three hours’
7. AI agents are stealing credentials in minutes, and security’s old response targets can’t keep up
9. Datacentres to create 70% more carbon emissions by 2030, IEA report warns
1. Nvidia privately confirmed it could track chips to fight smuggling, even as industry publicly questioned feasibility
This article looks at the US Chip Security Act following the alleged smuggling of thousands of Nvidia GPU chips to China.
- The Congressional Committee on Science, Space, and Technology is cosponsoring an act that would require AI chip manufacturers to introduce geo-tracking mechanisms into the chips.
- Nvidia has officially stated that geo-tracking is not technically possible. However, a senator on the congressional committee affirms that Nvidia has indicated to him, off the record, that geo-tracking is indeed possible.
- The technique uses a recent tool called Fleet Intelligence whose role is to remotely monitor the performance of GPUs running in data centers.
- The mechanism, however, only locates GPUs. An extension is being envisaged where a chip’s license could be remotely verified during the boot sequence. The boot could be interrupted if the license cannot be validated.
2. Jensen Huang’s departure blasts big hole in AI startup funding, VCs and founders say
TechCrunch examines the impact on AI startup funding by the resignation of Nvidia CEO Jensen Huang.
- Nvidia has been a prolific investor in AI startups. According to PitchBook, the company participated in 105 rounds already in 2026. For one analyst, “Nvidia is carrying almost every AI-native startup from application to neocloud”.
- In September, Nvidia acquired Hugging Face for 12.93 billion USD.
- Jensen Huang himself was seen as pivotal in these investments. Nvidia’s interim CEO, Colette Kress, is seen as far more conservative. This could impact the investment landscape. For one analyst: “When Nvidia was on the cap table, everyone felt safe”.
- Huang’s investment approach saw companies requiring between 100 million USD and 500 million USD in GPU chips raise half of that amount from Nvidia – in return for use of Nvidia chips. This has helped to solidify the Nvidia ecosystem.
- Less startup investment from Nvidia could open the way for other companies to sell their chips. Google for instance announced two new chips last week.
3. ‘Jensen Huang lost his mind’: how AI’s smartest man fell for the world’s oldest trick
The Guardian newspaper looks into what is a very old problem of industrial espionage in the Tech world following Jensen Huang’s resignation as Nvidia CEO.
- Espionage was already a problem during the Cold War in the 1970s and 1980s. A former FBI agent estimates that there were thousands of Soviet agents in the US trying to steal intellectual property secrets from Western firms.
- Currently, the FBI believes that a number of North Korean agents are active in the US with the intent of spreading malware in cryptocurrency development companies.
- The nonprofit Center for Strategic and International Studies has identified over 224 incidents of Chinese espionage in the US between 2000 and 2022, half of which aimed to steal technology secrets. Many of these cases use a “honeypot-style” approach where the suspected spy pursued personal or sexual relationships with the target.
- The general approach is for the spies to identify staff at companies with access to important information – this can be done by analyzing social media profiles. Staff with substance abuse or gambling problems are particularly vulnerable, as are employees seeking relationships.
- Protecting CEOs is particularly challenging since it requires lower-level staff to act as gate-keepers and to intervene. In the case of Nvidia, the current FBI investigation will seek to identify who in the company new about the affair with the alleged Chinese spy and whether the affair was allowed to happen despite knowledge.
4. How AI agents are changing the rules of software governance
This InfoWorld article looks at governance principles for software coding agents. As coding autonomy moves from the human to the AI, policy must no longer be simply a corporate document. Rather, policy must translate to engineering techniques that oversee agents.
- One technique is the creation of an audit trail. For instance, the agent environment should record each file and network access made by an agent.
- A second step is to maintain human accountability. In particular, a precise and meaningful human-in-the-loop policy needs to be defined. For instance, agents may be allowed to generate code, but any change to the database might need human approval.
- Another principle is to measure AI usage and cost. While, “tokenmaxxing” is used to measure and control token consumption, this does not capture the business value that the tokens are being spent on. A technique called a “software harness” is emerging where a software layer around agents carries business context so that agent activity value can be measured in business terms and controlled.
- An important lesson is that organizations are accountable for the actions and mistakes of their AI agents. AI Agents alone cannot give competitive advantage since bad actions can impact trust in the company. A trustworthy governance framework is required for competitive advantage.
5. Panic over cloud capacity is out of hand
This article considers the impact of misinterpreting headlines around the scarcity of compute power in data centers – and the impact this is having on organizational compute planning. In essence, capacity is currently tight as there is a high demand for GPUs, CPUs, and power supply. However, capacity is not yet sold out.
- Part of the issue stems from the cloud providers themselves who are communicating the narrative that they have demand for computing capacity that will last for years. This reassures their investors, but at the same time, confuses organizations into believing that future compute capacity is uncertain.
- The confusion is putting unnecessary pressure on enterprise architects who already have enough to deal with on data governance, model selection, integration, latency, and the shortage of qualified people.
- For AI, most organizations are building RAG infrastructures, fine-tuning smaller models, deploying inference workloads and testing copilots. The compute demands for these operations are far lighter than those required for training frontier models.
- The article suggests that organizations already segregate workload types based on capacity before compute availability planning. Workload types would include training, fine-tuning, inference, development, testing, batch processing, real-time applications, and data movement.
6. FBI says Artex AI open source hacking tool used to breach networks in ‘under three hours’
In the US, the FBI is warning that, Artex, the autonomous AI hacking tool is being used since July to launch autonomous-driven cyberattacks around the world. Targets include financial services, government, Tech companies, defense, phone and internet companies, and law firms.
- Artex is an open-source penetration testing system that can be connected to any large AI model to help probe network defenses. The tool allows penetration to move very fast.
- After initial detection of the use of Artex for cyberattacks, the company that owns the tool made it closed source. However, this step was taken too late and many copies of the tool still exist on the Internet.
- Further, the FBI indicates that copies of Artex and similar tools on the Internet have been hacked, which means that legitimate users of the tools are themselves vulnerable to cyber breaches.
7. AI agents are stealing credentials in minutes, and security’s old response targets can’t keep up
This VentureBeat article looks at how the increasing speed of cyberattacks due to the use of AI is forcing organizations to think about metrics used to describe attacks.
- In traditional human-led cyberattacks, the criminal launches perhaps dozens of commands in an hour. An AI agent orchestrating a cyberattack can launch thousands of commands.
- Breakout time is a cybersecurity metric introduced by security firm CrowdStrike in 2018 to represent the time it takes for a criminal who compromises a company machine to compromise a second machine on the network. In 2018, the average breakout time was 1 hour and 58 minutes; in 2026, the time is 29 minutes. The pertinence of the metric has become doubtful.
- The Palo Alto Networks’ Unit 42 organization writes that, in some cases, “attackers needed just 72 minutes to move from initial access to data exfiltration, 4x faster than last year”.
- The metrics currently used for attacking criminals are command throughput, which is the number of commands issued within a given period, and access-broker handoff, which is the time it takes for an intruder to pass access to a second criminal group.
- For defenders, metrics include mean time to defense, damage-to-discovery time, and agent revocation time. Mean time to defense measures how long it takes to plug a vulnerability. Damage-to-discovery time measures how long an attack goes on unnoticed. Agent revocation time measures how long it takes to remove agent permissions when a problem is discovered.
- Flashpoint research says the average time from public disclosure of a vulnerability to first exploitation was 115 days in 2024 and 44 days in 2025. The time available to patch vulnerabilities is shrinking.
- A challenge for cybersecurity teams is to present the problem to management in business terms. One CISO writes that a typical CISO “talks right past the CFO or the CEO because they focus on vanity metrics and they don’t phrase things in business terms”.
8. Amidst fears of AI cyberattacks, Satya Nadella says lots of critical infrastructure is ‘unpatchable’
Speaking at the recent New York Times’ DealBook Summit, Microsoft CEO Satya Nadella said that a lot of IT infrastructure in the US is running software that cannot be patched. This is a major concern in an era where AI is the principal means of discovering and exploiting software vulnerabilities.
- The smartification of ordinary devices – connecting standard devices to the Internet – is contributing to the problem. For Nadella, “there is a lot of stuff in our critical infrastructure that’s unpatchable” such as “firmware and the chillers (cooling device often used in data centers) and the HVAC (Heating, Ventilation, and Air Conditioning) systems everywhere else”.
- As the speed of vulnerability discovery has increased, so has the speed of patching systems. This leads to precipitation in patching, where the patch has not had the time to get thoroughly tested. This leads to problems when the patch is made.
- In the same interview, Nadella justified an offer made by Microsoft to 60’000 employees urging them to quit instead of being part of a huge layoff.
- He suggested that the reason was AI – not that AI was replacing their jobs, but because AI was transforming workflows so job descriptions are changing. For instance, he said “software engineering is very much alive and, in fact, is going to thrive, except the way the team is organized and the work that a software engineer does has now got a new operating point.”.
- Another example cited was Cowork, the AI agent in Microsoft 365 apps. This is developed by a team of 11 developers who have all “mastered” AI. The CEO said they “build a software factory that builds the product. They don’t build the product”.
9. Datacentres to create 70% more carbon emissions by 2030, IEA report warns
The International Energy Agency (IEA) has said that data centers will consume twice as much electricity by 2030 compared with today.
- Around the world, the next four years will see data centers generate 70% more carbon emissions as the US and China in particular use fossil fuels to power their data centers.
- Data centers will consume an extra 350 terawatt-hours per year – which is more than the electricity consumption in the whole of Great Britain.
- An agency report said that data centers contributed to the uncertainty about energy supply, in addition to the war in the Middle East, the continuing rise of air-conditioning, and the rising population.
- Today, data centers consume 120 gigawatts of electricity, with 80% being in China, the US and the EU.
- In the EU, data centers will use renewable sources of energy for about 70% of their electricity, the remaining electricity being used by nuclear power, gas generation, and a very small amount of coal power.
10. Anthropic on course to break $2 trillion barrier
Anthropic touched 2 trillion USD in market shares on its debut on the Nasdaq. The stock opened at 110.50 USD, 33% higher than the 83 USD IPO price.
- Anthropic CEO, Dario Amodei, and his six co-founders each own 10% of the company now have stakes worth 150 billion USD.
- The IPO has been seen as a “referendum” on AI. One analyst said: “I've looked at this whole week as a referendum on AI. Nvidia gave us a read on the buildout, and Anthropic is about to give us a read on investor appetite for the next generation of AI companies.”.
- Another cautioned that “some the demand may simply reflect investors not wanting to miss out on the rare opportunity to gain exposure to a leading AI company, even if its long-term profitability remains uncertain.”.
- Anthropic has nearly 900 billion USD in computing commitments with partners that include Amazon and Google. Its rapid growth depends on it being able to secure vast amounts of future computing capacity – which is not yet guaranteed.
- Despite the successful entry onto the Nasdaq, AI bubble concerns persist. Concern focuses on the circular deals among AI companies, chipmakers, cloud providers, and investors, which inflate the perceived growth in the sector.