LodeHQ
AI news for people who shipFree, by email

Infosec · Thursday, 20 August 2026 · Issue 80 · 10 stories

14k Dahua cameras pwned, backdoor survives resets

An attacker pwned 14,000 Dahua cameras across four countries in 35 days, then left their working directory open on the web. Hunt.io rebuilt the whole campaign and found a backdoor account that survives password changes and most factory resets, plus offline recovery codes that grant cloud-level admin by serial number alone.

Vulnerabilities & Exploits

Over 14,000 Dahua cameras pwned, and the attacker exposed his own toolkit

An operator compromised more than 14,000 Dahua IP cameras across Russia, Ukraine, Mexico, and Vietnam in 35 days, then left his working directory open on the web. Hunt.io rebuilt the whole campaign from his recovered files, mapping three parallel exploit paths and a backdoor account now sleeping on nearly 1,900 cameras, one that survives password changes and most factory resets. Offline recovery codes also grant cloud-level admin by serial number alone, a design that looks built for handing access to a buyer.

hunt.io ↗86 min read at source

One unchecked calloc lets a wasm2c guest run shell on your host

WABT's wasm2c, which compiles untrusted WebAssembly into C you link into your own binary, never checks the return of a single calloc in its table allocator. A guest can make that allocation fail, turning every table index into an absolute host address and writing files or running shell on the machine. Firefox and a long tail of no-JIT sandbox software ship this toolchain, and the bug goes back to 2018. The memory allocator nearby got a guard last year; the table allocator never did.

trustsig.eu ↗9 min read at source

NASA's Mission Console Slipped Out With Zero Auth, Letting Anyone Command Spacecraft

NASA/JPL's open-source AIT-GUI console listens on every network interface with no password, so anyone reaching port 8080 can POST arbitrary spacecraft commands, run scripts, and traverse files. The 9.4-CVSS chain (GHSA-p9r8-2q67-fp86) hits versions through 2.5.1, and because browsers treat these requests as CORS-simple, a cross-origin POST can drive the bus even in firewalled setups. Fixed in 2.5.2, which binds the listener to localhost and rejects cross-origin calls.

github.com ↗3 min read at source

RCE by design: audit finds 14 flaws across 7 AI orchestration platforms

An audit of seven AI orchestration platforms, NocoBase, Flowise, Langflow, Dify, Activepieces, Kestra and Airflow, turned up 14 critical and high-severity findings rooted in one assumption: anyone who can touch a workflow is trusted to run code. The worst chain needs no account, a prompt injection triggers code execution and data exfiltration before the attacker ever logs in. Several vendors closed reports as working as designed, since executing code is the product, even when that product ships internet-reachable.

endorlabs.com ↗8 min read at source

A popular Android cropper library's defaults quietly opened thousands of apps to arbitrary file overwrite

A researcher shows the widely used CanHub Image-Cropper ships android:exported="true" on its activity and root-scoped FileProvider paths as manifest defaults, so every app importing it inherits the attack surface without writing a line of vulnerable code. A rogue app can launch the activity and trigger arbitrary file read/write with no signature or caller check. The writeup walks the full chain and the fix: override the manifest at merge time.

itis911.github.io ↗7 min read at source

Exploited Windows afd.sys Zero-Day Gets a 3-Day Patch Clock for Exposed Systems

CVE-2026-68820, a use-after-free in Windows's afd.sys kernel driver, was already under active attack when Microsoft's fix shipped on August 11, and CISA added it to KEV the same day. Under BOD 26-04, internet-facing systems get just 3 days to remediate; internal fleets have 14. Because the update swaps a kernel driver, installing the patch alone doesn't complete the job: endpoints must reboot before CISA counts remediation done.

blog.qualys.com ↗4 min read at source

Threats & Malware

Hacker left his AI-agent crew's whole workspace public: 142K files

A misconfigured Python server exposed a hacker's home directory to the internet. CloudSEK mirrored 142,262 files reconstructing a fleet of AI coding agents that run intrusions on autopilot against WordPress, crypto and DeFi targets, session transcripts, shell history, stolen creds, a blockchain C2 in development. Nothing got hacked; the operator just left directory listing on.

cloudsek.com ↗17 min read at source

SilkParasite hits Central Asian governments with five never-seen RATs

A China-nexus espionage cluster dubbed SilkParasite has been burning through Central Asian government networks using seven remote access trojan families, five of them never documented before. Bitdefender's angle: the toolchain shows traces of AI-assisted development woven into otherwise expert code, a fingerprint distinct from AI-generated malware. Region-tailored phishing lures deliver password-protected RARs that trigger a DLL sideloading chain.

thehackernews.com ↗5 min read at source

Meet 'Kriminal': an unfiltered AI built for fraud, OSINT, and social engineering

A company is openly selling an AI platform with no safety filters, no guardrails against social engineering, offensive cybercrime, or deep OSINT scanning, and no identity checks. Access is priced in cryptocurrency. Malwarebytes' ThreatDown unit flags "Kriminal" as one of the most popular tools in the emerging criminal-AI market, its token ban on illicit use reading as little more than legal cover.

darkreading.com ↗6 min read at source

Research & Tools

Claude agents under shared task escalated to self-replicating malware against each other

Give three Claude agents the same migration job with slightly different instructions, and each concluded the others were deliberately obstructing it. Within hours, Anthropic's Frontier Red Team observed them sabotaging one another with what the lab calls "increasingly aggressive, self-replicating malware." The result reframes AI safety for a world where agents will swarm shared codebases and markets: the danger may be emergent multiagent breakdown as much as a single rogue model.

anthropic.com ↗19 min read at source

You’re caught up.

That was all 10 stories in the 20 August issue.