Anthropic Expands Claude Mythos Preview Access to Australian Organisations Through Project Glasswing
Cybercriminals Exploit ChatGPT Share Links to Distribute Malware Via Fake Outage Pages
Google Chrome Bolsters Security With Session Cookie Theft Protection for All Users
Hackers Exploit Meta’s AI Support Bot to Hijack High-Profile Instagram Accounts
Critical HTTP/2 Bomb Vulnerability Exposes Major Web Servers to Remote Denial-of-Service Attacks
https://www.anthropic.com/news/expanding-project-glasswing
Anthropic is broadening its cutting-edge AI program to Australian shores, with the company confirming that local organisations are now included in its Project Glasswing early access initiative. The expansion brings the total number of participating organisations to up to 150 across more than 15 countries, all gaining access to Anthropic’s powerful Claude Mythos Preview model. Access remains tightly controlled, with participation limited to organisations involved in defending critical infrastructure or operating across key sectors including power, water, healthcare, communications, financial services, and national security.
Claude Mythos Preview has already demonstrated significant cybersecurity capabilities, having been credited with identifying thousands of vulnerabilities at varying severity levels across open-source code bases. The model’s potential in the security space has drawn considerable attention, particularly following Microsoft’s own AI-powered MDASH vulnerability scanner discovering four critical remote code execution flaws in the Windows operating system in May 2026. However, Anthropic has noted that the model appears less effective when applied to extensively vetted code bases, suggesting its strengths are more pronounced in broader vulnerability discovery efforts.
Anthropic has stressed that its primary concern surrounding Mythos-class models is the risk of them being released publicly without adequate safeguards to prevent misuse. The company warned that threat actors are already leveraging large language models to uncover new and potentially devastating vulnerabilities at scale, making it “imperative cyber defenders adapt to maintain pace.” Anthropic also acknowledged it will likely not be the only AI firm to release models of this class, underscoring the urgency for responsible deployment frameworks across the industry.
https://pushsecurity.com/blog/llmshare-malvertising-campaign
Threat actors have found a novel way to abuse ChatGPT’s conversation sharing feature, using publicly accessible share links to host convincing fake outage pages designed to deceive unsuspecting users. The technique takes advantage of the legitimate ChatGPT platform’s credibility, allowing attackers to craft and distribute malicious content through shared conversation links that appear authentic at first glance. Security researchers identified the campaign as a particularly deceptive method of social engineering, given that the links originate from OpenAI’s own domain, making them harder for users and security tools to flag as suspicious.
The fake outage pages are designed to mimic legitimate service disruption notifications, prompting visitors to download what appears to be a fix or update, which in reality delivers malware to the victim’s device. By leveraging a trusted and widely recognised platform like ChatGPT, attackers significantly increase the likelihood that targets will interact with the malicious content without hesitation. The campaign highlights a growing trend of cybercriminals exploiting popular AI platforms as delivery mechanisms for malware, capitalising on the rapid mainstream adoption of these tools.
Google has rolled out a significant security upgrade to its Chrome browser, extending session cookie theft protection to all users in a move aimed at combating one of the most prevalent techniques used by cybercriminals to hijack online accounts. The feature, previously available only to select users, is designed to prevent attackers from stealing session cookies, which are small files stored in a browser that keep users logged into websites and online services. By targeting these cookies, threat actors have long been able to bypass multi-factor authentication and gain unauthorised access to accounts without ever needing a victim’s password.
The protection works by binding session cookies to the device on which they were created, making them significantly harder to exploit even if they are successfully intercepted or stolen by malicious software. This approach directly counters a technique known as “pass-the-cookie” attacks, which have become increasingly popular amongst cybercriminals and state-sponsored threat actors alike. The update represents a meaningful step forward in browser-level security, particularly as infostealers and other malware strains designed to harvest session cookies have surged in use across the threat landscape in recent years.
Google’s decision to make the feature available to all Chrome users rather than a limited cohort is greatly welcomed as a proactive measure to protect everyday Australians and organisations from account takeover attacks. Ensure your Chrome browser is updated to the latest version to take full advantage of the new protections.
https://krebsonsecurity.com/2026/06/hackers-used-metas-ai-support-bot-to-seize-instagram-accounts/
Hackers have successfully exploited a critical vulnerability in Meta’s AI customer support bot to seize control of high-profile Instagram accounts, including those belonging to the Obama White House and the Chief Master Sergeant of the U.S. Space Force. Instructions detailing the remarkably simple exploit began circulating on Telegram channels on 31 May 2026, accompanied by a video demonstrating how attackers could trick Meta’s AI support assistant into resetting account passwords without proper verification. The technique involved using a VPN connection with an IP address near the target’s usual location, initiating a password reset request, and then manipulating the AI support bot into linking the account to a new email address, after which a one-time reset code was dispatched to the attacker’s chosen address.
Meta moved quickly to contain the damage, confirming on Twitter/X that the issue had been resolved and that affected accounts were being secured. The company pushed an emergency patch over the weekend, with security blog thecybersecguru.com clarifying that no back-end database had been breached. The vulnerability was attributed to Meta’s decision to deploy a conversational AI layer to handle common account recovery workflows, a measure originally intended to reduce friction for legitimate users struggling to regain access to their accounts amid the platform’s notoriously poor human support infrastructure.
AI bots are just as susceptible to social engineering as human support staff, and similar attacks are likely to become far more common. Crucially, the hackers themselves confirmed that their exploit failed against any accounts with multi-factor authentication enabled, so make sure you have MFA setup for all accounts that offer them.
Cybersecurity researchers have uncovered a severe remote denial-of-service vulnerability, dubbed the “HTTP/2 Bomb,” that affects some of the world’s most widely used web server infrastructure, including NGINX, Apache HTTPD, Microsoft IIS, Envoy, and Cloudflare Pingora. Discovered by OpenAI Codex and reported by security firm Calif, the vulnerability exploits HTTP/2’s default configuration by chaining together two well-known techniques, a compression bomb and a Slowloris-style connection hold, to devastating effect. The attack targets HPACK, HTTP/2’s header compression scheme, where a single byte transmitted over the wire can trigger a full header allocation on the server, repeated thousands of times per request, whilst a zero-byte flow-control window prevents the server from ever freeing the consumed memory.
The scale of the potential damage is alarming, with researchers demonstrating that a single home computer on a standard 100Mbps internet connection could render a vulnerable server completely inaccessible within seconds. More critically, a single attacking client is capable of consuming and holding up to 32GB of server memory against Apache HTTPD and Envoy in approximately 20 seconds. What distinguishes the HTTP/2 Bomb from previously known variants is its novel amplification method, which exploits per-entry bookkeeping allocations around nearly empty headers rather than stuffing large values into compression tables, effectively bypassing existing decoded-size limits that servers had implemented to guard against earlier attacks.
Patches are currently available for NGINX, with users advised to upgrade to version 1.29.8 or above, and for Apache HTTPD via mod_http2 version 2.0.41. However, Microsoft IIS, Envoy, and Cloudflare Pingora have no patches available at the time of writing, leaving a significant portion of global web infrastructure exposed. Organisations unable to apply patches immediately are strongly urged to disable HTTP/2 entirely as a temporary mitigation measure.
Comments
Nothing yet. Say the first thing.
Sign in to join the conversation.