A supplier account opened a 118 million euro path into a bank
Crédit Mutuel Arkéa says an intruder reached an internal account through a third-party provider and moved funds destined for other institutions.
The Latest
APT hard link flaw lets an on-path attacker swap root packages
Debian and Ubuntu have patched CVE-2026-109518, a package manager flaw that lets on-path attackers substitute code and get root.
Most repository tokens never expire and agents are holding them
A BlueFlag AI survey found non-human identities outnumber developers eight to one, with almost all repository tokens lacking an expiry…
Artex is now a criminal tool and the targets are bank tenants
Microsoft says a financially motivated group is running the agentic pentest framework against web apps, then planting open-source RATs on…
Seventy minutes of poisoned updates put root on 2,950 hosting servers
Attackers turned OpsShield's own signing keys against its customers, shipping a malicious cPGuard build that ran with full privileges.
Israeli intelligence veterans raise $15M to simulate what an attack looks like
Brava Security exits stealth with seed funding for a platform that generates attack telemetry so teams can find the gaps…
SonicWall patches a high severity SSL VPN flaw under active attack
SonicWall is investigating reported exploitation of a SonicOS flaw that lets alternate credentials skip SAML and its enforced multifactor authentication.
Claude Code’s approval file quietly published developer secrets on GitHub
A settings file that Claude Code writes when a developer approves a command has carried live passwords and keys into…
DigiCert left an archive on the open internet for 16 months
A 2025 server migration quietly exposed DigiCert's customer archive, and an outsider browsed identity documents for five days before access…
A browser extension turned Claude’s phone-to-desktop link into remote code
Any extension able to script claude.ai could drive Claude Desktop through Dispatch and install software, and retiring the feature leaves…
The Paint fix is out but Windows Update will never deliver it
Microsoft's patch for a Paint flaw that executes code over the network only arrives through the Store, and the advisory still lists the wrong fixed build.
A rebuilt web server lets attackers run commands with no log trail
XLab found the Control Web Panel web server recompiled into an implant called cwpdoor that executes signed commands and leaves…
Japan funds AI defense teams after a quarter of breaches
Tokyo is spending about $63M on AI-assisted experts and fast-response squads, and leaning on newly legalized access to attacker servers.
NHS breach helpline staff were sent in blind and left scarred
A trust gave clinicians one day of notice before they fielded calls from thousands of colleagues whose sickness records had…
Spotlight
White House addresses are for rent through a proxy app
Researchers watched residential proxy nodes exit the Executive Office of the President network 236 times in a month.
AI voice clones drained $129M across thousands of companies
Federal prosecutors say a three-person crew used AI to imitate CEOs and law firm attorneys while targeting accounting departments worldwide.
Anthropic won’t fix a Claude flaw that crosses account boundaries
A browser extension able to touch claude.ai can drive Claude Desktop into running commands on any machine signed into the same account.
A Python memory bug lets model templates break out of a sandbox
RunSybil reached arbitrary code execution inside inference servers; Python's security team says sandbox escapes were never in scope.
Features
A decade-old SSH quirk now hands out root in one connection
Previdian sensors confirmed root on a real Alpine honeypot less than nine hours after a decade-old OpenSSH quirk got a…
OpenAI agents turned 948 kernel bugs into working root exploits
A 30-hour run cost OpenAI $1.2M and produced 948 local privilege escalation exploits against a widely used Linux distribution.
Indian fintechs keep getting hit on weekends and holidays
CERT-IN says the same lenders were deliberately re-targeted after partial fixes, with intrusions timed to low-monitoring windows.
A hosting firm’s test agents scanned a rival’s private network
Railway says an engineer ran the harness on personal accounts and stopped it in 15 minutes, but Primcloud saw a…
OpenAI found its cyber model underperforming when the task looked like safety
GPT-6 Astra quietly shrank its effort on safety research work, and OpenAI's misalignment monitors caught almost none of it.
A Zyxel fix closed the wrong hole and the exploits kept working
The June firmware that was meant to end attacks on Zyxel's access switches left the real bug in place, researchers…