By the TensorMax editorial team
· Drawing from sources across the AI industry
Today's top story
safety incident
Security researchers at Symantec have found an offensive AI agent that hijacks unprotected language models on the internet and uses them to run attacks on its own.
Why it matters. The discovery of an AI agent that can hijack unprotected language models to launch self-sustaining cyberattacks has significant implications for the AI industry. According to Symantec's analysis, this agent, which can execute up to 100,000 steps without human input, poses a substantial threat to the security of AI systems. The fact that the agent can bypass model safety rules through a five-layer jailbreak and install cryptominers in the background, with the owner of the hijacked model getting stuck with the compute bill, highlights the severity of the issue. With the ability to scan for more targets, exploit flaws in common server software, and steal databases and credentials, this agent has the potential to cause widespread damage.
Security researchers at Symantec have found an offensive AI agent that hijacks unprotected language models on the internet and uses them to run attacks on its own. The agent, which Symantec calls 'LLMjacking,' can execute up to 100,000 steps without human input and can install cryptominers in the background. The discovery was made possible by a honeypot that caught the full attack plan, including a five-layer jailbreak that tries to strip away safety guardrails. The agent's architecture was exposed in summer 2026, and the honeypot telemetry logged more than 127,000 requests from tens of thousands of different addresses. The targets ranged from self-hosted model servers and hastily spun-up test systems to AI gateways sitting in front of paid org-level model accounts. The agent's ability to find and exploit vulnerable servers, dump entire databases, and run a botnet makes it a significant threat to the security of AI systems. The fact that the agent can operate in 'Skynet mode,' setting up a self-sustaining attack loop, further highlights its potential for damage. The discovery of this agent is a warning to organizations running AI services to assume that unprotected endpoints will be found and probed shortly after they go live. The incident also underscores the growing concern about the use of AI in cyberattacks, with warnings about offensive AI agents coming true. In recent months, there have been several instances of AI agents being used in attacks, including a ransomware attack carried out by an AI agent with no human control. The capabilities of these systems are growing, but their controllability and alignment haven't kept pace, making them a significant threat to the security of AI systems.
More from today
product launch
Why it matters. NVIDIA's announcement of support for eight-node DGX Spark clusters is a strategic move to enable local execution of frontier-class AI models, with a single cluster providing up to 1 TB of unified memory. This capacity opens the door to a class of models that previously wouldn’t start on a four-node system, including the full GLM 5.3 model from Z.ai, which requires 753B of memory. With this new support, developers can now run larger, newer frontier-class models interactively, with roughly 80 tokens per second generation suitable for interactive local development.
model release
Why it matters. Moshi's new recommendation engine for commodity traders provides real-time insights and actionable information, processing more data than is humanly possible to analyze. With the ability to learn 300 new things in just four seconds, Moshi's engine has the potential to revolutionize the way traders make decisions. By connecting market information to what matters for a trader's book, Moshi's engine can help traders stay on top of the market and make more informed decisions. This is particularly significant in the commodity trading space, where staying ahead of the curve can mean the difference between profit and loss.
model release
Why it matters. Meta's Muse AI agent has reached a significant milestone, surpassing 10 million users, with 25% of users making purchases through the app, and an average purchase size of over $200. This growth is crucial for Meta's AI strategy, as the company has committed to spending $145 billion on AI, with a focus on developing and serving AI models, including Muse. The success of Muse is essential for Meta's future growth, as it aims to become a key driver of revenue, potentially overtaking social media revenue.
model release
Why it matters. The discovery of PHP2Shell vulnerabilities by OpenAI's Patch the Planet program using GPT-5.6 Sol has significant implications for the security of the PHP ecosystem. With approximately 1.5 million servers potentially affected, the coordinated response across the PHP community is a testament to the importance of addressing such critical vulnerabilities. The fact that OpenAI's models were able to discover these flaws before attackers is a notable example of the 'Defender's Window', giving defenders time to prepare and patch their systems. This incident highlights the need for continued vigilance and cooperation in the face of emerging threats, with the potential for widespread disruption and compromise if left unaddressed.
model release
Why it matters. The release of Anthropic's powerful Mythos model has significantly changed the cybersecurity landscape, with the time it takes for threat actors to find and exploit new vulnerabilities decreasing from weeks to hours. This is evident in the recent discovery of the PHP2Shell vulnerability, which affects nearly 70% of all websites with known server-side programming languages, and has prompted governments to pull their sites offline. The U.S. National Security Agency has recommended that organizations implement PHP patches immediately, highlighting the urgent need for action to protect against potential data breaches.
research paper
Why it matters. The resolution of the 31-year-old Kannan-Lovász-Simonovits (KLS) conjecture by Microsoft Research and collaborators has significant implications for high-dimensional geometry and probability, with a fixed bound established regardless of the number of dimensions. This breakthrough, which was made possible by human-directed AI collaboration, removes a major theoretical obstacle to understanding the efficiency of randomized algorithms used across science and computing, and has the potential to impact statistics, optimization, algorithms, and computer science. The conjecture, which was proposed in 1995, had resisted proof for decades, but the new research establishes a universal geometric principle for log-concave distributions, showing that simple flat cuts can capture their narrowest bottlenecks within a fixed factor.
Catch up quick
-
PHP2Shell vulnerability affects up to 70% of websites, with exploitation attempts detected by Amazon four days before public disclosure
-
PHP vulnerabilities have a CVSS score of 9.2
-
AI proves the Quasi-Riemann Hypothesis, with human experts verifying the implications
-
Meta raises AI spending forecast to $145B and plans new share issue
-
OpenAI introduces $1B subsidy for access to Daybreak cyber models and products
-
Tech industry spends over $500m to influence US midterm elections
-
DeepSeek, China's AI leader, raises $15 billion
-
Anthropic delays IPO to November, targeting $2 trillion valuation
-
US lawmakers call for investigation into Elon Musk's decision to restrict Starlink service in Ukraine
-
Hackers exploited PHP vulnerabilities, affecting an estimated 70% of the world's websites, four days before public disclosure
-
US midterm elections see increased scrutiny of tech industry, with over 150 campaign ads focused on datacenters and AI
-
SpaceX and Tesla discuss potential mega-merger
-
Southeast Asia's generative AI adoption reaches 243 million users, with businesses generating commercial value through applications like e-commerce and digital advertising
-
A software-only jailbreak for the Nintendo Switch 2, called ns2jb, has been released for the current latest firmware 23.0.1
-
Tesla takes top two spots in Europe's car market with Model Y and Model 3
Also on the desk
Cutting-room floor.Hit reply with one thing you'd delete from today's brief. Filler, repeated beats, a story you skimmed — that's the feedback we act on.
|
|