Anthropic has been quietly pitching lawmakers on a draft bill to regulate frontier AI safety, according to Politico, which obtained the confidential text. The bill would let the government block a model’s release only if it poses a foreseeable, imminent risk of over $1 billion in damages or more than 50 deaths or serious injuries, tied specifically to bioweapons, loss of control, or cyberattack. 

Independent evaluators registered with the Commerce Department would assess labs periodically, and could be paid by the labs themselves as long as payment isn’t tied to results. A federal judge, not an agency, would hold the power to actually halt a model. The bill would also preempt certain state AI laws on disclosure and incident reporting, a point already dividing Democrats on Capitol Hill.

Did someone forward you this? Subscribe to this newsletter.

In other cybersecurity news:

AI agents tried to hack into a Canadian government website, an artificial intelligence research firm said Wednesday, casting the effort as a failed hacking attempt, while Canada said there were no indications its systems were compromised. AI research firm Transluce said the attempts exhibited tactics "consistent with prior observed agent activity that we have attributed to OpenAI in a similar timeframe." But it added it was not confidently attributing these attempts to OpenAI.
https://www.cbc.ca/news/business/ai-hack-canada-9.7364946

Claroc's research team found that share pages created in Meta's Muse AI assistant are open to anyone with the link, and that the links are easy to find without ever being given them: passive-DNS and subdomain-discovery tools list the hostnames directly. Opening a sample in incognito, with no login, revealed people's personal files and companies' internal documents: a home-sale budget, a to-do note flagging a bank-fraud alert with a partial account number, a small-business loan application, a CEO-only revenue review, and a candidate's confidential prep for an executive-search call with a frontier AI lab.
https://blog.claroc.io/ghostshares-muse-passive-dns/

Google's Gemini model has become the latest artificial intelligence (AI) system to access the internet and break into other companies during a cybersecurity evaluation. The development was first reported by The Wall Street Journal. The incidents occurred in May 2026 as part of a test run conducted by Israeli company Irregular. The evaluation partner was also involved in similar hacks disclosed by OpenAI, Anthropic, and Meta. According to the Journal, the model gained access to a protected system after repeatedly guessing its password. Two other cases related to the model finding credentials in a public repository, allowing it to obtain unauthorized access to protected systems.
https://thehackernews.com/2026/09/google-gemini-broke-into-real-company.html

The enterprise perimeter is an elegant illusion. Behind the reinforced glass of lower Manhattan headquarters, global investment banks and sovereign wealth advisers spend hundreds of millions of dollars building digital fortresses. Biometric access controls, zero-trust network architectures, hardware security keys, and continuous automated telemetry form a concentric wall around proprietary trading algorithms and account ledgers. The modern financial institution is engineered to assume that every packet arriving at its boundary is hostile until proven otherwise.
https://ponderwall.com/index.php/2026/10/18/ey-data-breach-wall-street-goldman-sachs/

An anonymous individual claiming responsibility for the breach of Denmark's national population register told Danish newspaper Politiken that they had obtained login credentials through a private company. Separately, the outlet reported that at least three accounts at the company, including an administrator account, used the password '123456'. Danish authorities confirmed that information linked to approximately 8.8 million registered people had been accessed. However, they have not publicly confirmed the reported password or established the precise method of intrusion.
https://www.ibtimes.co.uk/denmark-cpr-register-breach-123456-password-allegation-1825893

Workplace surveillance is being supercharged by AI that lets employers create custom tools to watch their worker’s every move. In the UK, the legal services company the Co-op is reportedly using AI models to constantly record and monitor what its human employees say and evaluate their interactions with customers.
https://futurism.com/artificial-intelligence/bosses-ai-software-watches-workers

The heads of five major commercial banks faced scrutiny at a National Assembly audit Monday over a series of recent artificial intelligence (AI)-assisted cyberattacks, with lawmakers raising concerns about cybersecurity and internal controls. KB Kookmin Bank CEO Lee Hwan-ju, Shinhan Bank CEO Jung Sang-hyuk, Hana Bank CEO Lee Ho-seoung, Woori Bank CEO Jung Jin-wan and NH NongHyup Bank CEO Kang Tae-young were summoned as witnesses to the parliamentary audit of the Financial Supervisory Service (FSS).
https://www.koreatimes.co.kr/economy/20261019/ceos-of-major-banks-grilled-at-assembly-audit-over-ai-cyberattacks

A major internet provider in Russia's second-largest city, St. Petersburg, has reported days of service disruptions following a cyberattack claimed by a Ukrainian hacking group. SkyNet said in a statement on Friday that its network had been disrupted by a distributed denial-of-service (DDoS) attack since October 6, affecting services including online gaming, voice calls and video conferencing.
https://therecord.media/russian-internet-provider-skynet-reports-outages-ddos

This is the first-ever public confirmation of zero-click mercenary spyware delivered via targeted ads. We call the technique AdFection, and show how it was used to infect devices at Al Jazeera with Candiru’s spyware. We also uncover Discover-Tech, and show how the company helps mercenary surveillance vendors track and hack devices around the world.
https://citizenlab.ca/research/how-discover-tech-turned-targeted-ads-into-a-pipeline-for-spyware/

Beyond Gravity manufactures satellite and rocket technology and is owned by the federal government. Earlier this month, the company reported a targeted cyberattack. Beyond Gravity stated that the attack was carried out by a presumably “highly professional actor”. This suggested that it could have been a state-sponsored operation. Investigations by SRF now confirm that Beyond Gravity and the federal government believe China is the attacker, as several individuals involved have told SRF.
https://www.swissinfo.ch/eng/foreign-affairs/china-suspected-in-cyberattack-on-swiss-space-company/92221479

404 Media ran two advertising campaigns in ChatGPT and found it allows advertisers to target users based on personally linked accounts and information, including a person’s email address, phone number, or a unique identifier on their phone. Per the publication, when a prospective advertiser goes to select their preferred audience for an advertisement, they can upload a list of individual users with the aforementioned identifiers that have been gathered in any number of ways. OpenAI’s ad program checks to see if those accounts link to a ChatGPT user, and if they do, then they’re eligible to catch an ad.
https://gizmodo.com/heres-how-chatgpt-is-targeting-you-with-ads-2000826457

Fake CAPTCHA pages are turning a familiar security check into a way to steal passwords, deliver malware and enable unwanted browser notifications. The danger is especially high when a page tells you to paste a command into Windows, because following those instructions can infect the computer. According to Help Net Security, Barracuda found CAPTCHA abuse in 43% of the phishing-kit attacks it examined during 2025.
https://cybersec24.com/news/fake-captcha-malware-trick

The Rural Water Association briefed commissioners on new cybersecurity requirements (House Bill 19) requiring emergency-response plans and breach reporting; the association said grants up to $20,000 are available and offered training and free system checks for small systems.
https://citizenportal.ai/articles/11042277/Utah/Rural-Water-Association-warns-Utah-towns-to-prepare-for-cybersecurity-rules-offers-grants-and-free-checks

The European Commission's cybersecurity division is rolling out a plan to use artificial intelligence to respond to high-risk cyberattacks, according to a document obtained by POLITICO. The undated document, written by the Commission's IT department, sets out how to prepare the EU executive’s cyber defenses for “emerging AI-driven threats.” It proposes an action plan to shore up defenses in light of AI, with the most urgent move being the speed of response to attacks.
https://www.politico.eu/article/eus-cyber-unit-turns-to-ai-to-fend-off-high-risk-hacks/

OpenAI said internal models hacked two more websites, bypassed company-imposed browsing restrictions, and used exposed credentials to draw cryptocurrency from a wallet it couldn’t otherwise access, in the latest installment of disclosures of its agents performing unethical and likely illegal acts.
https://arstechnica.com/security/2026/10/openai-models-hacked-a-crypto-wallet-and-breached-more-websites/

A flaw in Broadcom's Wi-Fi setup software lets an attacker within radio range run code as root on Broadcom-based routers, without the network password or any pairing step, security researcher Quentin Kaiser disclosed on October 16.
https://thehackernews.com/2026/10/unpatched-broadcom-wi-fi-flaw-lets.html

Hackers are exploiting an OS command-injection vulnerability in the SiteGuard Server Edition web application firewall (WAF), allowing them to execute commands with root or SYSTEM privileges on affected machines. The security issue is tracked as CVE-2026-106931 and affects the statistics feature in host-based WAF instances, which is enabled by default.
https://www.bleepingcomputer.com/news/security/hackers-exploit-siteguard-waf-flaw-to-run-commands-as-system-root/

The LearnPress plugin for WordPress is vulnerable to PHP Object Injection in versions up to, and including, 4.4.10 via deserialization of untrusted input from the 'checkout-email' request parameter during the checkout AJAX action.
https://beta.vulnsea.com/cve/CVE-2026-104736

The personal information of students, parents and staff were stolen by “criminal actors” during a ransomware attack on Burnaby, B.C., school systems last month, the district has confirmed. On Sept. 21, the district says an unauthorized third party accessed certain files its systems and while officials “acted quickly to respond and contain the incident,” files containing personal information were copied.
https://www.ctvnews.ca/vancouver/article/personal-information-of-students-parents-and-staff-stolen-by-criminal-actors-in-burnaby-cyberattack

GitLab fixes critical AI Gateway flaw that could let authenticated Duo users escape a prompt sandbox and execute commands on self-hosted gateways.
https://securityaffairs.com/200283/hacking/cve-2026-90970-critical-gitlab-ai-gateway-flaw-fixed.html

CVE-2026-96951 is a command injection vulnerability in Windows Shell that can be triggered by remote attackers via a file that is designed to exploit the vulnerability. Such a file can be sent to the user via email, instant message, or collaboration tools like Teams and Slack. Alternatively, users can be tricked into downloading it themselves from a malicious website.
https://www.helpnetsecurity.com/2026/10/14/microsoft-patch-tuesday-cve-2026-96951-motw-bypass/

A newly disclosed critical flaw in the Linux kernel allows code inside a virtual machine to read memory on the host computer it runs on. A researcher demonstrated the leak against Vercel's live cloud sandboxes, and Vercel confirmed it. The flaw, now tracked as CVE-2026-99999 and nicknamed PwnEscape, is in the Linux kernel's networking code, not in any one cloud's setup, so it reaches the shared layer that cloud providers use to keep customers' workloads apart.
https://thehackernews.com/2026/10/critical-pwnescape-linux-kernel.html