The chart is interesting.

On the IPI benchmark, Opus 5 improved over Opus 4.8, reducing the probability of an attacker succeeding within 15 attempts from 5.5% to 2.0%, and from 0.5% to 0.2% on 1 attempt. It also improved on Sonnet 5 (5.9% at k=15) and Mythos 5 (2.6%), making it the most robust model evaluated. Opus 5 also outperformed all non-Claude models on this benchmark. The most robust non-Claude model was Muse Spark at 16.5% within 15 attempts—more than eight times Opus 5’s rate. The most capable GPT 5.6 variant, Sol, was comparable to its predecessor GPT 5.5 (20.0% versus 20.8% within 15 attempts), and was 10 times as likely to be successfully attacked as Claude Opus 5 at 2.0%. The other GPT 5.6 variants are less robust, at 30.4% (Terra) and 43.9% (Luna). A single attempt against GPT 5.6 Sol succeeded 3.1% of the time, higher than the 2.0% an attacker achieved against Opus 5 after fifteen attempts.

We know that preventing prompt injection is impossible in the general case. But we are getting much better at blocking it in specific cases.

This essay was written with Barath Raghavan, and originally appeared in The Guardian.

In July, Hugging Face, a company that hosts much of the world’s AI software and open-source AI models, was hacked. A malicious dataset had been used to run code on one of its servers. Whoever was behind it captured internal security credentials and moved through systems over a weekend, running thousands of actions from a swarm of temporary server environments. It looked like the work of a sophisticated criminal group.

It was not. It was one of OpenAI’s new, still unreleased GPT models.

Their science experiment had escaped the lab. OpenAI was running the unreleased AI model through a benchmark that tests how well AI can successfully hack systems. To push the limits and evaluate the AI’s true capability, the company switched off the safety filters that normally stop it from doing this kind of hacking. Aware that this could go wrong, they confined the AI to an isolated environment and denied it access to the internet.

But the new AI cheated. It took literally its goal to get as high of a score as possible. It broke out on to the open internet. It inferred, probably from its training data, that it could “solve” the task by getting the answers from Hugging Face’s servers. So it chained together stolen credentials and further unknown security exploits to hack the company’s network.

Nobody instructed the AI to do any of this. It was, in OpenAI’s words, “hyperfocused on finding a solution” to the test it was being given. And while this might seem like something new with AI, it’s really very old. This is how a genie behaves, and it is a key challenge with AI agents in general.

In folklore, genies—and other magical beings—grant wishes literally, not how the wisher intended. King Midas asked that everything he touched turn to gold, and starved. The sorcerer’s apprentice wanted the broom to fill the cistern, and it performed its task so well that it flooded the house.

We now have machines that do this. Ask a modern AI agent to save money on your phone plan and it might simply cancel the plan. Tell it to book a flight, and it might hack the airline website to override restrictions. Or, like OpenAI, ask it to do well on a test and it might break into another company to steal the answers. Each time, it recognizably completed the task you set, but it didn’t do what you would have wanted.

This isn’t malicious behavior. No one asked for, or wanted, Hugging Face to be hacked. OpenAI and Hugging Face and the AI were ostensibly on the same side, and the AI was trying to do what it had been asked. That’s what makes it so difficult to guard against: you can’t filter for bad instructions because the instructions were fine.

The gap is between the words we use and what we mean by them. We call that gap the Genie coefficient.

AI labs know this is a problem, and they’re quietly saying so. For example, the Chinese lab Moonshot recently warned that its latest AI model may have “excessive proactiveness” and “make unexpected decisions on the user’s behalf”. The UK’s AI Security Institute has started tracking “cheating behavior in frontier model evaluations”. We wouldn’t tolerate a car that is excessively proactive or ruthlessly efficient, and yet that’s the reality of AI today.

Improvement is possible. Just as AIs have gotten much better at resisting prompt injection attacks over the last few years, we can safely predict that they will get better at avoiding genie-like behavior. The point of the Genie coefficient is to track progress. AI companies like benchmarks, and they all work to compete to be the best.

Dozens of benchmarks and leaderboards tell us how well these AI models write code, perform logical reasoning, and pass standardized legal and medical exams. But there is nothing that scores whether a system does what you actually meant. We need to develop a measure for this, test it regularly, and push for improvement. We’re not going to have trustworthy AI agents without it.

Last week, national security agencies from the Five Eyes—that’s the rich, English-language-speaking countries club—jointly released a statement warning of the increasing cyber risks of AI models: in particular, their ability to autonomously hack into systems and networks. The statement was more measured than some of the breathless headlines about it, and the advice they gave is pretty much the standard advice everyone gives—albeit with newfound urgency.

Internet risks are nothing new, and cyberattacks—both large and small—have been a significant issue since long before the current crop of generative AI models.

What’s been changing over the decades, and what AI is changing even faster, is the gap between skill and ability. For most of human history, the two terms were synonymous—but computers have decoupled them. As the gap between the two expands, humans empowered with these AI tools can do more: more writing, more research, more analysis and also more damage than ever before. These models can, with little detailed direction, autonomously hack into networks, steal data, deploy ransomware and destroy systems. And to the extent there is a solution, it’s going to involve harnessing AI for the defense.

In 1998, seven people from the hacker group L0pht testified before Congress. They told a mostly clueless Senate committee that they could take down the internet in 30 minutes. That was partly real and partly bravado, but it illustrates an important point: hacking into systems, stealing data and causing damage all required skill.

Contrast the L0pht hackers with hackers derided as “script kiddies.” They didn’t understand computers, or security. Instead, they used hacker tools written by others. Their actions required minimal skill and even less knowledge. But once those hacking tools became widespread, the number of potential attackers increased.

That number has continued to increase, as quality and availability of prewritten attack tools has grown. And it is growing dramatically with AI. Today’s AI systems—not just the frontier models, but most of them—are capable of carrying out cyberattacks automatically. They all do better in the hands of skilled attackers, but increasingly they are able to act autonomously with only minimal prompting.

The thing about people with ability but no skill is that they are often outsiders, not part of any professional community, and not bound by any rules or norms. This phenomenon is much more general than in cybersecurity. Any doctor can tell you how to untraceably poison someone, and many virus researchers know how to create a bioweapon. Any bridge engineer can tell you how to place explosives to blow a bridge up. The reason that murderous doctors and terrorist engineers are so rare is that the lengthy process of acquiring those skills also instills a moral and ethical code. If every random person has access to good poisoning advice, that puts us all in danger.

Modern AI systems are, in effect, a universal adviser to help people do harmful things. And while the current AI megacorporations are trying to build guardrails to prevent people from asking questions whose answers will enable the questioner to do harm, that’s not going to work in the long term. Smaller, cheaper, open-source models, including models that can run on people’s computers, and especially groups of models that run in concert with each other, are just as good as the frontier models from companies like OpenAI and Anthropic. And they continue to get better. These models will be passed around from person to person, like script kiddie hacker tools, and they won’t have any such guardrails.

Instructing AI models to spy on people and report any malicious prompts to the authorities fails for similar reasons. The megacorporations can do that, but the locally run open source models won’t. This could buy us a few months at best.

A third possibility is to somehow make the models themselves unable to hack into computers, create bioweapons or do anything else that might harm people or society. That won’t work, for the same reason we can’t teach doctors how to treat poisonings without also teaching them how to poison. It’s the same knowledge. It’s the same with construction and demolition. And it’s the same with cybersecurity. We want these AI models to be able to review computer code, find vulnerabilities and automatically fix them. The benefit to our collective security will be enormous. Unfortunately, the same knowledge can be used for attacks.

Where this leaves us is in a world of increased volatility. Super-powered humans with AI assistants will be able to do both wonderful and horrible things.

This brings us back to the Five Eyes statement. Everything they recommend is something security professionals have been recommending for years, if not decades. They are things talked about at that congressional hearing back in 1998, titled “Weak computer security in government: Is the public at risk?” Even the Five Eyes admitted that their security advice is not new, only more urgent.

What’s new is how fast things are changing: “The rapid pace of frontier AI development means cyber risk assumptions can become outdated in months, not years. We must act before and be prepared to adapt and withstand evolving threats.” The Five Eyes point to AI technology—not necessarily chatbots, but AI more generally—being used to strengthen every aspect of defense, to “detect vulnerabilities earlier, improve software quality, monitor unusual behavior, and respond faster to incidents—reducing both the cost and impact of incidents.”

Excellent advice from the Five Eyes security agencies. We need to do this with every risk that AI heightens, not just cybersecurity.

This essay was originally published in The Guardian.

A new rowhammer attack gives complete control of NVIDIA CPUs.

On Thursday, two research teams, working independently of each other, demonstrated attacks against two cards from Nvidia’s Ampere generation that take GPU rowhammering into new—­and potentially much more consequential—­territory: GDDR bitflips that give adversaries full control of CPU memory, resulting in full system compromise of the host machine. For the attack to work, IOMMU memory management must be disabled, as is the default in BIOS settings.

“Our work shows that Rowhammer, which is well-studied on CPUs, is a serious threat on GPUs as well,” said Andrew Kwong, co-author of one of the papers. “GDDRHammer: Greatly Disturbing DRAM Rows­Cross-Component Rowhammer Attacks from Modern GPUs.” “With our work, we… show how an attacker can induce bit flips on the GPU to gain arbitrary read/write access to all of the CPU’s memory, resulting in complete compromise of the machine.”

Update Friday, April 3: On Friday, researchers unveiled a third Rowhammer attack that also demonstrates Rowhammer attacks on the RTX A6000 that achieves privilege escalation to a root shell. Unlike the previous two, the researchers said, it works even when IOMMU is enabled.

The second paper is GeForge: Hammering GDDR Memory to Forge GPU Page Tables for Fun and Profit:

…does largely the same thing, except that instead of exploiting the last-level page table, as GDDRHammer does, it manipulates the last-level page directory. It was able to induce 1,171 bitflips against the RTX 3060 and 202 bitflips against the RTX 6000.

GeForge, too, uses novel hammering patterns and memory massaging to corrupt GPU page table mappings in GDDR6 memory to acquire read and write access to the GPU memory space. From there, it acquires the same privileges over host CPU memory. The GeForge proof-of-concept exploit against the RTX 3060 concludes by opening a root shell window that allows the attacker to issue commands that run unfettered privileges on the host machine. The researchers said that both GDDRHammer and GeForge could do the same thing against the RTC 6000.

It’s called AirSnitch:

Unlike previous Wi-Fi attacks, AirSnitch exploits core features in Layers 1 and 2 and the failure to bind and synchronize a client across these and higher layers, other nodes, and other network names such as SSIDs (Service Set Identifiers). This cross-layer identity desynchronization is the key driver of AirSnitch attacks.

The most powerful such attack is a full, bidirectional machine-in-the-middle (MitM) attack, meaning the attacker can view and modify data before it makes its way to the intended recipient. The attacker can be on the same SSID, a separate one, or even a separate network segment tied to the same AP. It works against small Wi-Fi networks in both homes and offices and large networks in enterprises.

With the ability to intercept all link-layer traffic (that is, the traffic as it passes between Layers 1 and 2), an attacker can perform other attacks on higher layers. The most dire consequence occurs when an Internet connection isn’t encrypted­—something that Google recently estimated occurred when as much as 6 percent and 20 percent of pages loaded on Windows and Linux, respectively. In these cases, the attacker can view and modify all traffic in the clear and steal authentication cookies, passwords, payment card details, and any other sensitive data. Since many company intranets are sent in plaintext, traffic from them can also be intercepted.

Even when HTTPS is in place, an attacker can still intercept domain look-up traffic and use DNS cache poisoning to corrupt tables stored by the target’s operating system. The AirSnitch MitM also puts the attacker in the position to wage attacks against vulnerabilities that may not be patched. Attackers can also see the external IP addresses hosting webpages being visited and often correlate them with the precise URL.

Here’s the paper.

We don’t have many details:

President Donald Trump suggested Saturday that the U.S. used cyberattacks or other technical capabilities to cut power off in Caracas during strikes on the Venezuelan capital that led to the capture of Venezuelan President Nicolás Maduro.

If true, it would mark one of the most public uses of U.S. cyber power against another nation in recent memory. These operations are typically highly classified, and the U.S. is considered one of the most advanced nations in cyberspace operations globally.