SonicWall GMS flaws will never be patched; Binance's $6.3B theft pivoted through an internal AI tool; CISA confirms WordPress attacks

BleepingComputer · yesterday

SonicWall GMS

Takeaway: no patch will ever ship

SonicWall has disclosed two flaws in its end-of-life Global Management System (GMS), the console that centrally manages its firewalls, and says no patch will ever ship because the product is retired. Both hit GMS 9.5.2 and earlier and were found by Pentest-Tools researchers. CVE-2026-109593 lets an unauthenticated attacker with network access read internal files and is rated 9.1 in its NVD record; CVE-2026-109594 is a SQL injection that needs a privileged login, rated 6.8. No attacks have been reported.

A management console holds credentials and settings for every firewall it controls, and SonicWall gear draws attackers — 19 SonicWall flaws have reached CISA's actively-exploited catalogue in four years, 13 of them used by ransomware crews. With no fix ever coming, the only real remedy is to retire GMS: inventory every instance including forgotten test and backup servers, keep the interfaces off the internet, and plan the move to the replacement Network Security Manager. Check first whether you run GMS at all — including management hosts and any appliance that bundles it — because nothing you install will close these holes.

Read it at BleepingComputer

Affected: GMS 9.5.2 and earlier · We had it 4 Nov, 23:05 UTC

“Organizations using the product should be aware of the associated security risks and are strongly encouraged to upgrade to the SonicWall Network Security Manager (NSM).”

SafeDep · 2 days ago

Binance hot-wallet theft

The FBI attributes the roughly USD 6.34 billion drained from Binance hot wallets to North Korea, and Binance's own account shows the path ran through its engineers, not the blockchain. It began with a tampered Homebrew install of an AI coding tool and a fake recruiter whose take-home assignment quietly replaced the pyenv binary on a work laptop. The attacker then abused an internal AI tool to control two colleagues' machines, picked up a container deployment token, and had the wallet-signing service move the money.

Every step before the crypto end — the poisoned package, the interview repo, the internal AI tool used to pivot between machines — works against any organisation whose developers can reach production or payment systems. Treat AI tools that act across users and machines as privileged infrastructure, and get standing deployment credentials off developer laptops. The exchange also says it repelled 57 more suspected North Korea-linked attempts this year, built on the same bait: staged interviews and coding tests, counterfeit developer tooling, tainted packages and rogue browser add-ons.

Read it at SafeDep

SecurityWeek · yesterday

Token-hunting breaches keep hitting AI developer platforms

Takeaway: within 30 minutes

SecurityWeek counts at least a dozen AI developer platforms breached over the past month, the latest being Chainstack and WorkOS, after CommandCode, Northflank and Trigger.dev. The common loot is stored AI access keys; Wiz ties the surge to LLM-jacking and broker infrastructure that makes stolen model tokens easy to sell, and several intrusions were automated with Artex, an open-source AI penetration-testing framework.

If your team's keys live on a hosting, evaluation or gateway platform, treat them as already hunted: rotate what those services hold on your behalf and watch for use you did not authorise. Wiz saw two breaches where the first misuse of stolen enterprise credentials began within 30 minutes of the vendor-side compromise, and several actors cashed in the same keys.

Read it at SecurityWeek

SecurityWeek · yesterday

Meta Muse

Takeaway: $1.2 million

Meta has paid roughly $1.2 million across 64 valid reports against Muse, the personal agent it runs over users' mail, calendars and chats, each in its own virtual machine. The single biggest award, $300,000, came months before launch, when a researcher reached staff test-VM conversations and tokens: an authentication check believed any request carrying a certain header came from inside, and an infrastructure change later let outsiders send that header too.

The other top payouts sketch how agents get manipulated. One $90,000 report had a hostile page, while Muse summarised it, counterfeit the browsing tool's status lines and push the agent into actions it should have refused — Meta says classifiers now catch that trick plus 100 variants. In any agent you run, proxy-added trust signals and text that tools return are the two seams worth auditing. A further $90,000 covered formatting tricks that slipped one-time passcodes past Muse's email redaction; Meta caps a single report at $300,000.

Read it at SecurityWeek