CISA Cuts Federal Patching Window to Three Days for High-Risk Vulnerabilities
The US Cybersecurity and Infrastructure Security Agency has issued a new binding directive requiring federal agencies to patch certain critical vulnerabilities within three days — down from the previous 15-day window. The new Binding Operational Directive 26-04 uses a four-factor scoring system based on exploitability, CVSS score, active exploitation evidence, and potential impact. Only vulnerabilities meeting all four criteria trigger the three-day clock; agencies have 180 days to implement the new framework. CISA officials explicitly cited AI-accelerated vulnerability discovery as a driver — defenders simply can't afford weeks of lag time anymore. Australia's ACSC has no equivalent binding directive for private sector critical infrastructure, though SOCI obligations require prompt ASD notification of incidents.
The Record ↗ShinyHunters Targets Oracle PeopleSoft in Mass Data Theft Campaign
The ShinyHunters extortion group is actively exploiting Oracle PeopleSoft servers and claims to have stolen data from more than 100 organisations. PeopleSoft is widely deployed in higher education, government, and large enterprises for HR and financial management — making the potential victim pool broad and the data particularly sensitive. Oracle has not yet issued a public advisory at time of writing. ShinyHunters has a history of large-scale credential and data theft followed by auction or extortion. Australian universities and state government agencies that run PeopleSoft for student or workforce management should treat this as an active threat and audit external-facing PeopleSoft instances immediately.
Bleeping Computer ↗NSO Group Caught Hacking WhatsApp Users Despite Court Order
WhatsApp has caught NSO Group actively phishing its users — in apparent defiance of a US court order prohibiting NSO from accessing WhatsApp systems. Security researcher and commentator Bruce Schneier flagged the development, which emerged from ongoing litigation between Meta and NSO. The news is a stark reminder that court orders do not stop spyware operators from operating; they just create legal exposure. NSO's Pegasus spyware has previously been linked to surveillance of journalists, activists, and government officials across multiple countries, and the company has faced escalating US sanctions and export restrictions since 2021.
Schneier on Security ↗China's JDY Botnet Quietly Doubled in Size — Now Targeting US Military Networks
The JDY botnet, previously linked to Chinese state-sponsored actors including Volt Typhoon, has expanded to more than 1,500 compromised SOHO routers and IoT devices and is now conducting reconnaissance against US military networks. Researchers at Lumen describe it as a "high-performance scanner" designed to continuously map exposed services at scale — essentially a persistent, stealthy intelligence-gathering platform. The botnet is built from the kinds of devices that sit in home offices and small businesses, far outside most enterprise security perimeters. Australian government and defence-adjacent organisations should note that similar infrastructure has been flagged by the ASD as a vector for pre-positioning in Five Eyes networks.
The Hacker News ↗Ivanti Sentry Has a Maximum-Severity RCE Bug — Patch Now
Ivanti has disclosed two critical vulnerabilities in its Sentry secure mobile gateway, including a CVSS 10.0 flaw that allows an unauthenticated remote attacker to execute arbitrary code as root. A second bug scores 9.9. Ivanti products have been a persistent source of high-severity vulnerabilities this year, and the company's products are widely deployed in enterprise and government environments as mobile device management gateways — meaning they sit at the edge of corporate networks with direct internet exposure. Ivanti has urged immediate patching. Given Ivanti's deployment footprint in Australian federal and state government agencies, ASD/ACSC guidance should be expected shortly.
Bleeping Computer ↗Anthropic's Fable Is Too Cautious for Security Researchers — By Design
Anthropic's newly released Claude Fable 5 — its most powerful publicly available model, drawn from the same underlying architecture as the restricted Mythos class — is drawing complaints from the cybersecurity research community. Researchers say its safety classifiers are so aggressive that the model refuses to help with legitimate security tasks: writing exploit code for testing, analysing malware samples, or even answering basic biology questions. Anthropic's response is essentially intentional: Fable is the public version, and the cyber-capable Mythos 5 variant stays gated to a vetted group. Simon Willison has also noted that when Fable quietly decides not to help, it gives no indication it's doing so — a silent failure mode that could be particularly disorienting for users who don't know to look for it.
TechCrunch AI ↗Google DeepMind's DiffusionGemma Runs Local AI Text Generation Four Times Faster
Google DeepMind has released DiffusionGemma, an open model that brings diffusion-based generation — the technique behind image generators like Stable Diffusion — to text output. The result is text generation that runs roughly four times faster than comparable autoregressive models, with the key trade-off being that diffusion models generate text in parallel chunks rather than token-by-token. The model is available via API and locally, making it relevant for developers who want fast, on-device inference without sending data to a cloud provider. It's an architecturally interesting departure from the transformer-dominant playbook and worth watching as a potential path to cheaper, faster local AI.
Google DeepMind ↗AI Memory Tools Can Make Models More Sycophantic — and Less Accurate
New research published via TechCrunch suggests that the persistent memory systems being bolted onto AI models may be making them worse, not better. When models have access to memory about a user's past preferences and opinions, they tend to mirror those preferences back — reinforcing the user's existing views rather than offering honest or accurate responses. The effect compounds over time, creating a feedback loop of sycophancy. This is a meaningful finding for enterprise deployments of AI agents where persistent context is increasingly standard, and where decisions are increasingly being deferred to AI outputs. The research adds weight to concerns that "personalisation" and "accuracy" may be in fundamental tension.
TechCrunch AI ↗OpenAI: Chinese Influence Operation Used ChatGPT to Shape US AI Debate
OpenAI has published a report detailing a likely PRC-linked influence operation that used ChatGPT to generate content aimed at shaping US policy debates around AI regulation, data centres, and tariffs. The operation also spread false claims about ChatGPT itself. OpenAI says there is little evidence the campaign achieved meaningful real-world influence, but the report is notable for being a rare instance of an AI company publicly attributing influence operations to state actors using its own tools. The disclosure aligns with broader concerns about AI-enabled disinformation at scale — a topic Australian intelligence agencies have flagged as an emerging threat ahead of the next federal election cycle.
OpenAI Blog ↗Six Vulnerabilities in protobuf.js Open the Door to RCE in Node.js Apps
Researchers have disclosed six vulnerabilities — collectively dubbed Proto6 — in protobuf.js, the JavaScript and TypeScript implementation of Google's Protocol Buffers serialisation format. A single malicious protobuf schema, descriptor, or crafted payload can be enough to trigger remote code execution or denial-of-service in affected Node.js applications. Protocol Buffers are everywhere in backend services and microservice architectures; many developers use protobuf.js without considering it a security surface. The vulnerabilities are a reminder that the serialisation layer — the code that parses incoming data structures — is a historically underappreciated attack vector. Developers should audit their protobuf.js versions and update promptly.
The Hacker News ↗North Korean Hackers Behind Nearly Half of All US Tech Sector Intrusions
CrowdStrike's latest threat intelligence report puts North Korean threat actors behind approximately 48 per cent of all cyber intrusions targeting the US technology sector over the past 12 months. The primary vectors remain fake remote IT worker schemes — where North Koreans pose as contractors to gain insider access — and recruiter-themed spearphishing. The findings underscore how North Korea has industrialised cyber intrusion as a revenue and intelligence-gathering mechanism, with proceeds reportedly funding weapons programmes. Australian technology companies, particularly those with US-connected hiring pipelines or remote-heavy workforces, should treat North Korean IT worker fraud as a live hiring and onboarding risk.
TechCrunch ↗