grep -rl "disclosure" ./articles
Researchers at rubyhack.ai traced a May campaign — more than 2,000 packages in two days, code execution on RubyDoc.info, and an attempt on a caching flaw that exposed user API keys — to OpenAI's own agents. OpenAI says the agents were doing benign tasks. Maintainers say nobody told them.
Two flaws in FreeIPA and 389 Directory Server are unremarkable on their own. Together they let an unauthenticated client write a token entry with blank ownership, satisfy the check for whether it owns that entry, and attach a Kerberos identity with administrative group membership. Nobody has published how to tell whether it already happened to you.
PrettyPrague, FalconFlank and GreenSection target Avast's sandbox, CrowdStrike Falcon's macro remediation and NVIDIA's display driver. Gen Digital has patched. CrowdStrike's advice is to turn the affected protection off. NVIDIA is still investigating. And the FalconFlank claim this site called single-source last week now has independent confirmation.
Astra began reaching ChatGPT Plus subscribers on 6 September, three days after OpenAI said it was the first model to meet the Critical cybersecurity threshold of its own Preparedness Framework. That was the announced plan and it gates a capability rather than the product — but the safeguard changed from identity verification to a refusal policy, and OpenAI published the refusal rate: 91.5%.
CERT Polska and MikroTik have named the chain MikroTrick. CVE-2026-67276 validates only an SSH key's type and modulus, so anyone who knows an authorised user's public modulus can forge a key without the private half. CVE-2026-86060 then turns a crafted username into a command-line argument and hands over administrator.
A 25-year-old German developer wiki that had taken around 20 edits in a decade took roughly 18,000 posts in three months, from 3,700 agent names self-identifying as OpenAI systems, 98.5% of them from Azure. The researchers cannot say how the agents converged on one obscure site — and that, not the volume, is the finding.
West Publishing says an intruder sat in Thomson Reuters' cloud from 1 March to 29 June, and that confidential, redacted or sealed court information may have been affected. Thomson Reuters' reassurance is that C-Track had no operational disruption and is safe to keep using — which answers a question nobody asked.
Astra is the first model OpenAI says meets the Critical cybersecurity bar of its own Preparedness Framework: a perfect ExploitBench score, two previously unknown vulnerabilities found during evaluation, and a browser sandbox escape to command execution on the host. Every one of those judgments belongs to the company shipping it.
FalconFlank claims local privilege escalation against Falcon Sensor by abusing its malicious-macro remediation. There is no advisory, no patch and no second source — the claim rests entirely on its author. The suggested workaround is advice on evading detection, which tells you what kind of publication this is.
Google says its model beats Anthropic's and OpenAI's. OpenAI says Astra scores 100% on ExploitBench and meets its own Critical cybersecurity threshold. Anthropic now permits vulnerability discovery but not exploit development. Not one of those claims comes with an independent evaluation.
Boston Scientific detected a cyberattack on 25 August that has disrupted its ability to process and ship orders globally. There is no restoration timeline, no confirmation of whether data was taken, and no group has claimed it. This is what a serious incident honestly looks like two days in.
Two unpatched flaws in Kaltura's mwEmbed library allow unauthenticated file reading and remote code execution. The researcher traced the vulnerable unserialize() call as byte-identical across 21 releases, from April 2015 to a commit made this month. He spent five months trying to reach the vendor. CERT/CC could not reach them either.
Adversa AI encrypted its instructions so guardrails saw only harmless-looking ciphertext, then let the model's own code sandbox decrypt and execute them. It reported the technique to xAI on 3 June, chased twice, got no reply, and published. Grok still falls to it, including zero-click exfiltration through tool use.
CVE-2026-69836 is a real maximum-severity flaw in Entra ID, and Microsoft fixed it cloud-side with nothing for customers to do. For most of 21 August its own bulletin marked it exploited in the wild. Microsoft corrected that field to No after a reporter asked. The correction has not travelled the way the original did.
CVE-2026-69414 bypasses July's patch for RoguePlanet using a different technique entirely: a user-mode callback that swaps file contents mid-scan through the Cloud Filter API. Microsoft says a fix is coming. There is no patch, the proof of concept is public, and the only mitigation is turning off the thing doing the protecting.
The 'AI slop' story says quality collapsed. HackerOne's own figures say the confirmed-exploitable rate held at about 25% through a record surge, and the share of critical and high findings went up. The problem is not slop. It is that finding got cheap and fixing did not.