New tools, products, platforms, funding rounds, and company developments in AI security.
Researchers at the University of Toronto created an AI-powered computer worm prototype using only free, small language models (LLMs, which are AI systems trained on large amounts of text) that could self-replicate across a simulated network by finding and exploiting vulnerabilities (security weaknesses) and misconfigurations. The research shows that attackers don't need cutting-edge AI models to launch widespread network attacks, since using paid models would create detection points where safety filters could block malicious prompts.
Perplexity AI's CEO announced the company plans to hold an initial public offering (IPO, where a private company sells shares to become publicly traded) in 2028 regardless of whether competitors Anthropic and OpenAI succeed with their own IPOs. The CEO acknowledged that major IPOs from SpaceX, Anthropic, and OpenAI will test investor demand for expensive tech company offerings, but expressed confidence these companies deserve high valuations because they lead in AI model capabilities.
The Hades Campaign is a sophisticated malware attack targeting Python developer environments that uses multiple advanced techniques: it harvests credentials, replicates itself across systems, extracts sensitive data from computer memory, and uses adversarial prompt injection (tricking AI security scanners by hiding malicious instructions in plain text) to evade detection by AI-powered security tools. The malware enters through compromised Python packages and uses the Bun toolkit (a JavaScript runtime) to execute payloads while bypassing traditional security controls.
OpenAI has filed confidential paperwork with the US Securities and Exchange Commission to pursue an initial public offering (IPO, a process where a private company sells shares to the public on the stock market) at some point in the future, though the company has not decided on timing yet. This move intensifies competition with rival AI company Anthropic, which announced similar IPO plans one week earlier, as both companies compete for users, customers, and investors with valuations approaching $1 trillion. Going public would provide these AI companies with billions of dollars in capital, which they need because running AI systems requires enormous compute costs (the infrastructure and processing power needed to build, train, and operate AI models).
OpenAI has filed confidentially to go public on the US stock market, with an expected valuation exceeding $850 billion, making it one of the largest IPO (initial public offering, when a private company first sells shares to the public) listings in history. The company announced the filing preemptively because it expected the confidential submission to become public anyway, and stated it has not yet decided on a timeline for going public.
OpenAI has confidentially filed for an IPO (initial public offering, where a private company sells shares to the public for the first time) with the Securities and Exchange Commission, joining competitors Anthropic and SpaceX in preparing to go public. The company, valued at over $850 billion, has not decided on a timeline for the IPO but says it could happen as soon as late 2025, though leadership notes some goals may be easier to achieve while remaining private. OpenAI is also planning a tender offer to let employees sell shares at the current valuation to ease pressure for immediate liquidity.
OpenAI has filed a confidential Form S-1 with the SEC (Securities and Exchange Commission, the US agency that oversees financial markets and public companies), which is a preliminary step toward becoming a publicly traded company. This move follows Anthropic's similar filing in June, as both AI companies compete to go public. The confidential filing keeps sensitive details like executive pay and financial information private for now, rather than making them immediately public.
Apple announced new AI features at its developer conference, including an improved Siri that can have back-and-forth conversations with users. Unlike some competitors, Apple is partnering with Google and Nvidia to build its most advanced AI model (Apple Foundation Model Cloud Pro) while emphasizing privacy by keeping user data local and using a system orchestrator (a routing component that directs queries to either on-device or cloud processing) to minimize data collection compared to AI services like ChatGPT.
Apple announced a major upgrade to Siri at its developer conference, integrating it with Apple Intelligence (Apple's AI tool) and rebranding it as 'Siri AI'. The new voice assistant will function more like AI chatbots (such as ChatGPT or Google Gemini) rather than a traditional web-search question-and-answer tool, with a wide release planned for fall.
A critical vulnerability was found in Zcash's Orchard privacy pool, a system that uses zero-knowledge proofs (mathematical techniques that prove something is true without revealing details) to hide transaction information. The bug allowed attackers to bypass input validation checks and create fake ZEC (Zcash's cryptocurrency) out of nothing, though it's now been fixed.
Google is upgrading NotebookLM, an AI-powered note-taking app, to use Gemini 3.5, a newer and more advanced version of its AI model that will provide more accurate answers. The update lets users start research projects by simply asking questions, and NotebookLM will automatically search the web to find relevant sources rather than requiring users to manually import materials.
This item is not AI/LLM-related and does not describe a technical security issue, vulnerability, or problem. It is a corporate legal announcement about a confidential S-1 filing (a registration statement for going public) with the SEC (Securities and Exchange Commission), noting that the company expects the document to leak and is making a preemptive announcement while deciding on timing for a public offering.
Microsoft's AI chief Mustafa Suleyman discusses how Microsoft has restructured its AI division to independently pursue superintelligence (AI systems that could surpass human capabilities across all domains), following a renegotiated partnership with OpenAI in October that allows both companies to develop models separately. The interview covers Microsoft's new approach to training frontier models (cutting-edge AI systems at the limits of current technology), the company's relationship with OpenAI, and how AI is being perceived by the public and in politics.
This article presents 15 strategic questions that CISOs (chief information security officers, the leaders responsible for an organization's security) should regularly ask themselves to ensure their security programs stay effective and aligned with business needs. The questions cover areas like demonstrating security's value to the business, protecting critical processes, understanding impact of breaches, detection speed, and keeping pace with modern threats. The article emphasizes that security programs must continuously adapt rather than remain static, especially in an AI-enabled threat environment where the focus should shift from finding every vulnerability to protecting the most critical business processes and ensuring rapid incident response.
OpenAI introduced Lockdown Mode to reduce data exfiltration (unauthorized theft of data), a security feature that disables external capabilities like web browsing and file downloads. However, security experts say the mode only partially limits data theft and doesn't fully block it, especially since attackers could find alternate paths (prompt injection, or tricking an AI by hiding instructions in its input) to steal data.
Fix: Lockdown Mode can be activated within OpenAI products' settings and limits web browsing to cached content, limits image support, disables Deep Research and Agent Mode, denies users the ability to approve Canvas-generated code to access the network, and prevents ChatGPT from downloading files for data analysis, though it can still operate on manually uploaded files. Alternatively, security professionals can implement isolation through their own enterprise controls such as network segmentation, least privilege access, Zero Trust concepts, application controls, and air-gapping (physically isolating networks).
CSO OnlineEngland and Wales plans to trial AI legal assistants in crown courts to reduce case backlogs, but lawyers warn the technology should not replace funding and staff. Concerns have been raised about AI hallucinations (false information generated by AI systems), including cases where AI created fake legal citations that were used in court decisions, highlighting risks to the justice system's integrity.
This AWS security digest from May 2026 covers updates across AI security, infrastructure protection, identity management, and compliance. Key topics include securing agentic AI workflows (AI systems that take actions autonomously) using frameworks like Cedar for authorization, protecting AWS networks with filtering tools like AWS Network Firewall and WAF (web application firewall, which blocks malicious web traffic), and establishing compliance programs for ISO standards. The resources include blog posts, code samples, and workshops to help organizations adopt these security practices.
This newsletter covers multiple AI and tech developments, including OpenAI's plans to transform ChatGPT into a 'super app' (an all-in-one application combining multiple tools and services) before going public, Google's $30 billion deal with SpaceX for AI computing power, and concerns about AI's rising energy costs and environmental impact. It also reports on facial recognition tools being deployed by immigration enforcement, fears about 'recursive self-improvement' (AI systems automatically improving their own capabilities), and how machine learning is helping historians analyze historical records while introducing risks of bias and errors.
Anthropic launched Project Glasswing in April to help companies find software vulnerabilities (weaknesses that attackers can exploit) using their AI model, though claims about its superiority over other models are unverified. A status report shows the project is finding many vulnerabilities, including dangerous ones, but almost none have been patched, and Anthropic has not released detailed information about the findings.
Most enterprise security teams are unprepared for real cyberattacks because they treat cybersecurity as a compliance requirement rather than an operational capability that requires constant practice. The military achieves rapid, coordinated responses to cyber incidents through regular, realistic exercises and by assuming attacks are inevitable, while businesses rely on outdated annual tabletop exercises and focus on prevention rather than detection, containment, and recovery.