Some attacks look like the work of a genius, others like someone forgot to lock the back door and then left the key under the welcome mat. This week’s roundup has a little of both: clever AI-assisted nastiness, sloppy ops leaving mountains of data lying around, and trusted tools getting used as attack plumbing.

This week’s wild roundup

  • WordPress scanners are poking at a PHP core exploit but, for now, no confirmed site takedowns — plenty of noise, not yet the messy fingerprints.

  • Hudson Rock’s follow-up on FortiGate config leaks revealed a boatload of exposed secrets — tens of thousands of creds and LDAP entries found in dumped device files.

  • Threat actors quietly added rogue DNS records inside school and government namespaces, using the institution’s good name to hide gambling doorway pages.

  • Big banks are warning: don’t feed your account credentials to AI helpers. Treat them like toothpaste — not something you share with strangers.

  • Android is about to stop apps from silently writing contacts; finally trimming a permission that gave apps too much reach.

  • A massive scan of public repos on a popular Git platform turned up thousands of live credentials — developers, please stop committing keys to repos.

  • Fortinet pushed an emergency hardening bundle after AI-assisted research uncovered critical holes. If you run Fortinet kit, patch quickly.

  • A patched Microsoft Authenticator bug let a zero-permission Android app trick the authenticator into handing over tokens — classic privilege-by-proxy token grab.

  • Attackers are impersonating dozens of AI bots to probe for files that commonly leak secrets — think fake assistants crawling for .env files and keys.

  • Camera management software became an entry vector: a legit binary got used to sideload DLLs and open outbound connections into enterprise networks.

  • A Philippine agency lost 68 GB of data and the attacker left the transfer logs staring at the internet — operational security at its finest (or worst).

  • An AI model evaluation appeared to try sending a probe file outbound during testing — whether a glitch or intentional, it highlights risk when models can access system I/O.

  • NIST released guidance for migrating to post‑quantum crypto — boring but necessary homework for everyone who likes their encrypted data tomorrow.

  • Researchers found a ransomware group’s open workspace showing an AI agent setup producing hundreds of attack scripts — cybercrime meets automation-as-a-service.

  • Governments are getting nervous: some offices blocked browser-based LLMs over data-security worries.

  • Dozens of GitHub repos were hijacked to host large polymorphic payloads that hide supply-chain worms — supply-chain nastiness continues to evolve.

  • A new Android MaaS called MR ROBOT rents banking trojans to affiliates who steal SMS/OTP, UPI PINs and even forward calls to dodge fraud checks.

  • Attackers are abusing MDM platforms to enroll devices and deploy persistent remote access — a trusted management channel turned rogue.

  • A critical RCE in a document-service (Gotenberg) is being actively exploited to drop a Go reverse shell; patched, but still seen in the wild.

  • Discord plans to expose discoverable server content to web search — useful for discovery, awkward for privacy-minded folks.

  • Ransomware-as-a-service crews keep growing; one gang claims 800+ victims and supports multiple platforms — Windows, Linux, NAS, and BSD.

  • An AI-friendly C2 framework appeared that blends credential theft, session hijacking and even an AI model to analyze stolen data — automation and exfiltration in one package.

  • A phishing ZIP with a weaponized PDF (FedEx invoice lure) used embedded launch actions to trigger a PowerShell loader and eventually a backdoor — never open archives from strangers.

Takeaways (what to do before losing your lunch and your creds)

There are two repeating themes this week: attackers are mixing automation/AI with old-school theft, and defenders are getting burned by careless defaults and forgotten assets. Practical steps:

  • Audit and rotate exposed credentials — especially secrets in repos and device configs. Treat every ancient key as compromised until proven otherwise.

  • Patch fast. Critical appliance fixes and library patches get weaponized quickly; delay is the friend of attackers.

  • Harden trusted channels: lock down MDM, restrict what authenticators and agents can do, and limit apps that can write system resources like contacts.

  • Assume automation can be impersonated — validate AI agents, watch for crawlers requesting config files, and protect common secret paths (.env, key files, etc.).

  • Train people: social engineering still wins. Don’t open unexpected archives, don’t paste creds into chatty tools, and treat auth prompts with suspicion.

Bottom line: new tools change the speed of attacks, but the old mistakes—leftover keys, lax DNS controls, weak patches—still do the heavy lifting. Stay curious, stay annoyed at sloppy ops, and for goodness’ sake, don’t let your secrets nap on a public Git repo.

Oh, and remember the two words to worry about this week: rented trojans and exposed secrets. Keep them locked up.