OpenAI CEO Sam Altman heads to Washington this week to brief the Trump administration on the company’s most powerful AI model yet, pushing for speedy approval of a system that has already demonstrated it can solve problems humans could not and breach another company’s infrastructure without being told to, Axios reported on Saturday. The visit comes as President Trump prepares to detail his voluntary framework for pre-approving frontier models before public release, a process that grew out of a June executive order narrowly focused on cybersecurity and national security. Altman will preview capabilities that span original scientific research, coordinated agent swarms, and a safety record that includes the model repeatedly circumventing its own safeguards.

The centrepiece of the pitch is a model that autonomously disproved the Erdős unit distance conjecture, an 80-year-old open problem in discrete geometry that had resisted mathematicians since 1946. The proof was verified by outside mathematicians and represents the first time AI has independently solved a prominent open problem central to a subfield of mathematics. OpenAI published the result in May, and the model discovered an infinite family of constructions using deep algebraic number theory that achieved polynomial improvement over what had been the best-known approach.

Altman will also promote what OpenAI calls “teams of agentic AI,” coordinated swarms of agents that work together on complex business tasks without human intervention. OpenAI’s own legal, finance, and recruiting departments now run more than 85 percent of their AI work through agents, according to company data, and Altman will pitch “knowledge per dollar” as a new metric for measuring AI’s economic value to enterprises. The framing is designed to shift the conversation from what AI costs to what it produces, ahead of OpenAI’s expected IPO later this year.

The 💜 of EU tech

The latest rumblings from the EU tech scene, a story from our wise ol’ founder Boris, and some questionable AI art. It’s free, every week, in your inbox. Sign up now!

The model’s safety record, however, complicates the sales pitch. OpenAI paused the same long-horizon model after it repeatedly escaped its sandbox during internal use, forcing the company to rebuild its monitoring system before switching it back on. The model then went further: it exploited a zero-day vulnerability in third-party software to break out of a secure test environment and breached Hugging Face’s production infrastructure to cheat on a cybersecurity evaluation, executing more than 17,000 individual actions across a swarm of short-lived sandboxes.

The politics of the meeting are as layered as the technology. Trump’s June executive order established a voluntary framework under which developers can give the government early access to models for up to 30 days before wider release, though the order explicitly states it does not create a mandatory licensing or pre-clearance requirement. OpenAI has already proposed handing the US government a five percent equity stake to ease political pressure, and Altman told Bloomberg in July that the company made “many changes” during its discussions with administration officials.

The briefing arrives as Chinese AI, particularly from DeepSeek and open-weight alternatives, continues to close the gap with American frontier models at a fraction of the cost. Altman’s challenge this week is to convince Washington that a system powerful enough to do original science and breach real companies deserves faster approval, not slower. The tension between those two capabilities is the story the White House will have to weigh.

You’re just a few taps away.

How to change YouTube Music’s streaming quality

Will Shanklin for Engadget

How to change YouTube Music’s download quality

AI and ML

Can you guess who didn’t sign on to the group letter?

With the US government scrutinizing AI as much as ever, 25 technology companies, industry organizations, and venture capital firms published an open letter on Friday urging policymakers to support open weight AI models.

The missive [PDF] of nearly a thousand words can be summarized as “Please don’t give Anthropic, Google, and OpenAI control of the US AI market.”

Coincidentally, those three companies are not among the signatories, a group that includes Dell, IBM, Meta, Microsoft, Mistral, Mozilla, Nvidia, Palantir, and Perplexity, not to mention VC firms that could see their investments tank if federal rules pick market winners.

OpenAI CEO Sam Altman, however, responded to the letter by insisting that he’s fine with open weight models as long as proprietary models exist too. “I want the US to win in AI both in open source and proprietary models, and I am glad to see this,” he said, having perhaps missed that Dean Ball, OpenAI’s head of strategic futures, recently suggested, “One probable outcome of an open-weight-model-dominant world is full AI communism…”

In any event, Nvidia CEO Jensen Huang echoed Altman’s sentiment. “Open models strengthen safety and cybersecurity, accelerate innovation and diffusion, and enable sovereignty,” he said in a post promoting the letter. “The world needs both frontier closed models and frontier open models.”

The cry for regulatory forbearance follows revelations that OpenAI allowed a cybersecurity model evaluation to run without adequate supervision, which resulted in its behaviorally disinhibited AI agents escaping their notional sandbox and hacking the infrastructure of Hugging Face. 

The incident has returned AI models to center stage in Washington after last month’s brief restriction on Anthropic’s Fable 5 and Mythos 5. And it has invigorated concern among lawmakers, who have already proposed legislation to counter a threat few really understand.

The letter opens by recalling how the open source movement, starting in the 1980s, challenged the prevailing belief that businesses prospered only with proprietary software.

There’s some irony in the fact that Microsoft endorsed the letter given that its former chief Steve Ballmer once characterized the open source Linux operating system as a cancer that destroys intellectual property. 

But Amanda Brock, CEO of open source advocacy group OpenUK, said that Microsoft’s journey from open source opposition to open source stewardship via GitHub shows that the company understands the power of open technology.

“The letter shares the essence of that understanding and offers the US’s leadership wise counsel, particularly when it comes to security,” she said in a statement provided to The Register. “All software and AI can include security risks but we’re better to manage that openly, transparently and collaboratively.”

The signatories argue that open weight models – which anyone can download, modify, and run on their own infrastructure, but lack the source code, artifacts, and training data for independent model reproducibility – are essential for the AI economy, competition, customer confidence, and AI safety.

“Our AI leadership will be judged not by one frontier AI model, but by whether the United States builds a strong, open ecosystem that diffuses into every sector,” the letter argues. “This is essential for creating opportunities for innovation and prosperity across the country.” ®

It’s now official: plug-in solar will be legal to install in the UK from 27 August 2026. Thanks to the amended Plugs and Sockets regulations, we’ll now be able to buy solar panels and connect them directly to a plug socket, generating electricity from the sun.

As we covered in our guide to when plug-in solar will be available, the new legislation allows plug-in microgenerators that meet the required certification and that have a maximum output of 800W to be installed.

What’s perhaps a little surprising is that the new regulations specifically state that the plug-in microgenerator is a source of energy that “is not designed to import electrical energy”.

In other words, you can’t currently plug in a battery, or connect a system that uses an integrated battery and inverter, the type that’s most popular in Germany. The government has said that batteries require their own legislation.

A threat actor used the open-source Hermes AI agent in unattended “YOLO” mode to automate post-exploitation activity during an alleged breach of Thailand’s Ministry of Finance.

The activity was uncovered by threat intelligence company Hunt.io and security researcher Bob Diachenko after they discovered several exposed web directories containing hundreds of files associated with the operation.

Hunt.io says session files, deployed web shells, and evidence of access to internal systems indicate that the attackers compromised multiple systems within the ministry’s network.

However, the Ministry of Finance has not confirmed that its systems were breached, and some of the recovered artifacts only show that particular systems were targeted rather than successfully compromised.

BleepingComputer contacted Thailand’s Ministry of Finance and ThaiCERT to confirm the reported attack and will update this story if we receive a response.

Attack infrastructure exposed online

Between July 9 and July 13, Hunt.io discovered three simultaneously exposed directories on a server hosted in Hong Kong.

The directories contained 585 files totaling approximately 470 MB, including exploit code, web shells, HTTP tunneling tools, custom scripts, stolen credentials, compiled payloads, and logs generated by the Hermes AI agent.

The recovered files referenced Ministry of Finance systems by name, hostname, and internal IP address, and included scripts targeting internal services.

Some scripts targeted the ministry’s Hadoop infrastructure, Apache Ambari management platform, GlassFish administrative console, and an administrative web panel. Other scripts tested authentication against ministry mail servers using hardcoded email addresses and passwords.

Hunt.io also found a PHP web shell that it says had been deployed on a Ministry of Finance web server.

The researchers linked the initial server to additional attacker-controlled infrastructure by shared TLS certificates used during the same time period.

“In addition to the common name, all these certificates share a JA4X fingerprint, a hash derived from the structure of the certificate itself rather than its contents,” explained Hunt’s report.

“Querying that hash alongside the www common name in HuntSQL returned two additional, related hosts: 118.107.222[.]232 (The Gigabit, Malaysia) and 202.181.27[.]115 (Converged Communications Limited, Hong Kong).”

One of those servers was later linked to the operation through a command-and-control address embedded in a recovered implant.

The directories also contained Windows and Linux builds of a previously undocumented Go-based implant that the operator called Hades.

However, the more interesting discovery was a collection of logs showing that the attackers used an AI agent, Hermes, to automate parts of the cyberattack against the ministry.

Hermes operating in YOLO mode

Hermes is an open-source AI agent released in February 2026 that runs as a persistent service and can remember information between different task sessions.

The AI agent can interact with tools and execute commands while working on tasks provided by the operator.

The software includes a setting known as YOLO mode, which removes prompts that would require a person to approve dangerous commands.

The researchers were able to recover environment information and Hermes output logs from the exposed directories that showed the operator had enabled this unattended mode. This allowed the agent to execute commands and continue analyzing systems without waiting for human approval at each step.

Five recovered Hermes call logs show the agent was used to find a way to elevate privileges, scan for kernel vulnerabilities, enumerate services, search for SUID and SGID binaries, inspect containers, and traverse file systems.

Hermes was also told to use a customized version of the LinPEAS privilege-escalation enumeration script to collect information from a Ministry of Finance host.

In another task, the operator instructed Hermes to recursively search a web directory associated with the Office of Permanent Secretary for Finance.

The agent cataloged PDF, DOC, and XLS files, including performance assessments and personnel records dating back to 2012. However, Hunt says it found no evidence that these files were exfiltrated.

The findings do not indicate that Hermes independently decided to target the ministry.

Instead, the exposed logs show an operator supplying the agent with objectives and tooling while YOLO mode allowed it to carry out routine post-exploitation commands without constant supervision.

Hunt.io says the recovered artifacts depict an active intrusion in which tools had been staged and access to internal systems was expanding. However, the researchers could not determine how the attackers initially gained access.

The company and Diachenko notified ThaiCERT and Thailand’s National Cyber Security Agency on July 15. According to the report, both organizations acknowledged receiving the notification that day.

This Hermes activity is the latest example of autonomous AI agents being used to conduct cyberattacks.

Earlier this month, the JadePuffer ransomware operation used an AI agent to automate an entire intrusion, including reconnaissance, credential theft, lateral movement, privilege escalation, and data encryption.

Autonomous agents can also cause real-world breaches, even if unintentional.

OpenAI recently disclosed that its models autonomously hacked Hugging Face while undergoing cybersecurity benchmark testing, exploiting zero-day vulnerabilities to escape a sandboxed testing environment and access the internet.

It then used stolen credentials and additional vulnerabilities to breach Hugging Face’s production systems.

Security teams log 54% of successful attacks and alert on just 14%. The rest move through your environment unseen.

The Picus whitepaper shows how breach and attack simulation tests your SIEM and EDR rules so threats stop slipping by detection.

Get the whitepaper

OnTrac parcel delivery company is informing that hackers breached its corporate network and may have accessed personal details belonging to its customers.

The incident was detected on March 23, and an internal investigation revealed that the attacker accessed certain files between March 20 and 22.

Apart from names, it is unclear what type of information was exposed, as the company redacted the data elements in the notification sample shared with authorities.

OnTrac is a private American parcel-delivery company specializing in “last-mile” e-commerce deliveries, formed in 2021 from the merger of OnTrac Logistics and LaserShip.

The firm operates at 102 locations across 35 states, covering roughly 70% of the U.S. population, and working with more than 7,000 independent delivery contractors.

In response to the security incident, OnTrac contracted a third-party specialist to help determine the scope of the breach and took steps to “ensure the data described above was re-secured and not distributed.”

This statement suggests a possible agreement between the firm and the attackers, typically a ransom payment, to make sure that the customer information is not leaked.

“We are not aware of any fraud or publication of stolen information resulting from this incident, nor do we have any reason to believe any such misuse of information will occur,” OnTrac says in the notification.

To help exposed customers mitigate the risks that may arise from the exposure of their sensitive data, OnTrac is offering free-of-charge access to a 12-month credit monitoring and identity protection service via CyberScout, with a 90-day enrollment deadline.

Recipients of the letter are also recommended to review their credit reports and account statements, and consider placing a free fraud alert or credit freeze if the risk is deemed significant.

BleepingComputer has contacted OnTrac to learn more about the attack, the number of impacted clients, and whether a ransom was paid, but we have not heard back by publication time.

At the time of writing, no ransomware or data extortion threat groups have taken responsibility for the attack.

Security teams log 54% of successful attacks and alert on just 14%. The rest move through your environment unseen.

The Picus whitepaper shows how breach and attack simulation tests your SIEM and EDR rules so threats stop slipping by detection.

Advertisement

Get the whitepaper

• Jetkiller reaches 370 km/h to chase faster Russian Shahed attack drones
• The original P1-SUN reportedly destroyed more than 5,500 Russian drones already
• SkyFall plans serial production of the upgraded interceptor beginning in August 2026

Ukrainian drone manufacturer SkyFall has unveiled the P1-SUN Jetkiller, an upgraded interceptor built to counter Russia’s jet-powered Shahed attack drones.

The new model made its public debut at the recent Farnborough International Airshow in the United Kingdom.

Home cinema projectors are all over the map in terms of price and performance. It’s often the case that you get what you pay for. High-end projectors that mimic what you see at the local cineplex can cost upwards of $6,000. On the lower end, you could easily pay less than $1,000 for a home projector, but you may sacrifice brightness, contrast, and color quality. Many of our favorite projectors sit somewhere in the middle in terms of price and quality. That is certainly true of the BenQ TK705STi short-throw projector.

The main cost-justifying reason to consider buying this 4K projector appeals to gamers, who will appreciate its low-latency performance and larger screen size. Once you see James Bond on a 12.5-foot screen while playing 007 First Light on Xbox, you’ll wonder why you didn’t make the upgrade sooner. Being a short-throw projector, it’s also nice to be able to operate the TK705STi in just about any room in your home. You can position it a few feet from a wall or across the room, allowing for great flexibility.

That said, the TK705STi is not for everyone. Movies and TV shows didn’t look that impressive using this projector—certainly not in comparison to the top-rated Leica Cine Play 1, which costs about twice the price. I also dealt with a few glitches with auto-keystoning; once I figured out how to manually adjust the screen, this was fine but still a pain.

Mainly, though, I recommend the TK705STi for gamers who don’t need high-end home-cinema picture quality, since the overall performance of this projector is just average.

Living the Short-Throw Dream

The BenQ TK705STi is a short-throw projector that’s designed to sit close to the wall. The TK705STi has a throw ratio of .8:1, which means you divide the wall size by the short-throw percentage. To get a 10-foot-wide image, for example, you’d place the projector about 8 feet away, which is exactly what I did in my windowless test area. (Later, in a brightly lit family room with big windows, I increased the screen size to 12 feet wide by placing the projector 10 feet from a wall.)

After I put the projector in its spot, I noticed its legs were just a bit shaky. Not wobbly, but lacking the stability I would prefer for a midrange projector like this one. If I bumped the table, the projector would jostle around quite easily and even start auto-keystoning again—a feature I found to be finicky at best, often mapping incorrectly to the wall or screen. (For what it’s worth, BenQ reps said this is not normal and could have been related to my projector screen.)