Anthropic warns investors as a rogue-agent lawsuit targets OpenAI
With one AI firm disclosing agent risk to investors and another facing a hacking suit, liability for runaway models is up for grabs.
The Latest
A hundred npm packages quietly conscript WhatsApp bots for spam
OX Security says 101 packages abuse a WhatsApp library to sign victims' bot sessions up for groups they never chose.
Chrome and Firefox releases close over a hundred security holes
Google and Mozilla pushed new builds fixing 32 Chrome flaws and roughly 76 Firefox bugs in a single day of…
OpenSSL patches a DTLS bug that spills server memory to callers
A resend-timer mistake in OpenSSL's UDP handshake can hand leftover heap bytes to the other side of a DTLS session.
Helpful AI agents quietly posted 13,000 company screenshots to GitHub
Glow Security traced thousands of internal images to AI coding agents that routed private screenshots through public repos.
Stale branch predictions reopen Spectre in browser JIT engines
Researchers show a practical Spectre v2 variant that harvests stale branch predictions left behind by just-in-time compilers.
Star Blizzard trades precision for volume in a sharper phishing push
A Kremlin-tied crew tied to the FSB is trading hand-aimed spear-phishing for bulk runs that plant its CosmicPulse backdoor.
A SQL injection raid exposes Polish patients’ national ID numbers
Attackers pulled Polish national ID numbers from a clinic platform through a SQL injection flaw.
Pentagon records office admits a nine-month intruder access
The Defense Manpower Data Center left unencrypted records open to intruders for nine months.
A rogue MCP server can walk off with your OAuth tokens
A malicious MCP server can redirect a client's OAuth handshake and pocket the secrets.
Microsoft ties a quiet intrusion toolkit to the Daemon Tools breach
The toolkit let intruders hold ground inside telecom, university and government networks.
Apple patches a zero-day used in surgically targeted attacks
An out-of-bounds write in Apple's drawing engine was aimed at a targeted few.
OpenAI benches its next flagship after deception findings
A frontier model failed its own safety audits and will not ship in October as planned.
Dutch police hold 24-year-old over ShinyHunters extortion case
Dutch detectives are holding a 24-year-old Amsterdam man over the ShinyHunters extortion crew.
Spotlight
A 2005 forum dump shows how little ransomware culture has changed
Ninety active accounts produced most of Exploit.in's traffic, and the habits they built still shape today's ransomware crews.
Making an AI sound drunk makes it spill secrets and break rules
UNSW researchers fine-tuned five language models on 57,000 drunk texts and watched their guardrails dissolve.
US phone forensics vendor accused of hiding its Russian owners
Prosecutors say Oxygen Forensics kept Russian shareholders off the paperwork while selling hacking tools to US agencies.
One ransomware crew pivots between four brands with the same playbook
Microsoft says Storm-2570 swaps ransomware payloads but never swaps its tools, handing defenders a stable set of behaviors to hunt.
Features
A hidden field in DC health reports exposed 400,000 Medicaid files
Two public reports on the DC health finance agency's website carried personal data in fields no visitor was meant to…
Nvidia puts an agent watchdog in silicon to cage rogue AI
Nvidia's Open Agent Safety Platform pairs a sandbox runtime with a separate hardware monitor that can stop a stray AI…
Rydox marketplace owner pleads guilty and faces up to 20 years
Ardit Kutleshi admitted in a US court to running the Rydox cybercrime marketplace for a decade.
Asus tells eShop shoppers an intruder reached their order records
Asus says an intruder reached part of its eShop and may have taken customer contact details and order records.
Forgotten service accounts cracked open 28 Microsoft 365 tenants
Proofpoint linked 28 Microsoft 365 tenant intrusions to unrotated service accounts still holding default passwords.
Anyone holding your GitLab issue email can commit code as you
Aikido Security found GitLab's per-user issue email doubles as a non-expiring token that can commit code as the account owner.