Threat Intel & Field Notes
Coverage that doesn't reduce to a single vendor advisory: infostealer and RAT write-ups, threat-actor campaigns and infrastructure takedowns, tooling roundups, and industry analysis on where security practice is falling behind.
The signature was there. The trust wasn't.
DigiCert's EV certs, WebEx and Zoom installers, ViPNet's signed updater. Three subverted trust chains this week, one design assumption behind them.
Mythos at three months: measure exposure, not volume
Three months after Anthropic's Mythos disclosure, the industry is still arguing about CVE queue depth. The number that matters is time-to-patch on your exposed critical assets.
AI-agent sandboxes are only as tight as the host tools
Pillar walked the same escape out of Cursor, Codex, Gemini CLI, and Antigravity in one week. The pattern isn't new — the trusted host tool is.
Ostium's LP vault down $23.75M after oracle-feed forgery
Attackers compromised off-chain price signing for Ostium's Arbitrum perpetuals DEX, submitted forged price attestations, and drained $23.75M from the LP vault.
Cursor, Codex, Gemini CLI, Antigravity: sandbox escapes
Pillar Security walks the same file out of the sandbox in four AI coding agents — each time by getting a trusted host tool to run what the agent wrote.
HollowGraph hides M365 C2 in calendar events dated 2050
Group-IB's HollowGraph hides M365 command-and-control in calendar events dated 2050-05-13, moving tasking and stolen files through legitimate Graph API traffic.
Exposed WebDAV lab: 1,048 artifacts, real Mexico victims
Rapid7 found an exposed WebDAV server with 1,048 attacker artifacts — QA'd lures, three tested CVEs, and 2,384 confirmed launch hits against Mexican targets.
Trend Micro: 'bandcampro' ran botnet ops through Gemini CLI
Trend Micro forensicated 200 Google Gemini CLI sessions used by a lone Russian-speaking actor to run an eight-node dental-clinic botnet through natural-language prompts.
Hugging Face confirms breach by autonomous AI agent
Hugging Face disclosed unauthorized access to internal datasets and service credentials by an autonomous agent framework that ran thousands of sandboxed actions across a weekend.
nginx patches heap overflow in worker (CVE-2026-42533)
F5 shipped nginx 1.30.4/1.31.3 and NGINX Plus 37.0.3.1 for CVE-2026-42533, a worker heap overflow reachable when a map directive uses regex capture variables in a string expression.
CERT-UA: UAC-0145 (Sandworm) runs ClickFix on Ukraine
CERT-UA alert 6318437 attributes a June–July ClickFix campaign hitting at least 10 compromised Ukrainian sites to UAC-0145, a Sandworm sub-cluster tied to GRU.
Kaspersky details HelloNet abuse of ViPNet updater
Kaspersky says an unknown APT — low-confidence Chinese ties — has abused the InfoTeCS ViPNet update client to plant Russian orgs since May.
Metasploit adds HTTP-to-SMB NTLM relay, RISC-V payloads
Rapid7's July 17 Metasploit wrap-up ships a Windows HTTP-to-SMB NTLM relay module, RISC-V shell payloads, and 421 new fetch-style variants. Check SMB signing tonight.
Microsoft ties ACR Stealer surge to WebDAV, blockchain C2
Microsoft's July 16 writeup links a late-April through mid-June ACR Stealer surge to WebDAV-hosted payloads and a blockchain dead-drop for C2 updates.
Two indicted over $43M laundered from investment scams
DOJ charged two New York-based Chinese nationals with laundering $43M in investment-fraud proceeds through 140 bank accounts and roughly 45 shell companies.
HollowByte: 11-byte OpenSSL DoS, no CVE, silent June fix
Okta's Red Team named 'HollowByte' — an OpenSSL DoS where 11 bytes of TLS pull 131 KB of process memory per shot. OpenSSL patched it in June with no CVE.
Agent Data Injection: The Bug Under Every AI Agent
Seoul National / UIUC / Largosoft research shows web and coding agents get steered by planted content in the pages, comments, and reviews they consume. Fix the trust boundary, not the model.
Flare finds carders still hunting clean IPs post-NetNut
Flare's read of 2,889 underground posts finds carders scrambling for 'clean' residential IPs two weeks after the FBI's NetNut seizure disrupted supply.
EY discloses breach via third-party IT ticket system
Ernst & Young says an unauthorized party accessed a third-party support ticket platform used by its IT staff between March 28 and April 12. Detection followed on April 23; disclosure landed July 17.
Armenia detains Aleksandr Ermakov on US REvil warrant
Russian tourist Aleksandr Ermakov has been held in Yerevan since 2026-06-28 on a US extradition request for a REvil suspect of the same name. His lawyer says the paperwork carries no patronymic.
OtterCookie's fake interview now steals AI-tool configs
Elastic Security Labs catches the DPRK's Contagious Interview crew hiding a four-stage payload in SVG country flag files — and the new file stealer specifically hunts .claude, .cursor, .gemini, and .windsurf configs.
GoSerpent: Go RAT hits APAC gov, TetrisPhantom overlap
Kaspersky documents GoSerpent, a Go-based RAT hitting Southeast Asian government and diplomatic entities since late 2025. Operational overlap with TetrisPhantom.
ACR Stealer, ClickFix, and why the Run box still works
Microsoft's Defender Experts detailed two ACR Stealer chains Thursday. Both start with a Run-dialog paste — and walk out with browser tokens and M365 files.
The plumbing behind $43M in investment-fraud losses
DOJ charges two in a New York-based network that laundered at least $43 million from pig-butchering-style investment scams through ~140 accounts.
UAT-11795 hides Starland RAT in trojanized installers
Cisco Talos names UAT-11795 — a financially motivated Russian actor pushing Starland RAT and bespoke WLDR C2 via trojanized WebEx, Zoom, MobaXterm installers.
AI can find the bug. Proving it is still the job.
SANS Fellow Stephen Sims argues the noise-to-signal ratio in bug bounty has shifted, but the proof-of-exploit standard hasn't — Bugcrowd's own policy shift agrees.
Elastic: TELEPUZ ClickFix stealer confirmed since April
Elastic Security Labs pins TELEPUZ, a modular C stealer spreading via ClickFix since late April, likely MaaS, with a Go Vidar variant as stage two.
Agent Data Injection: SQL injection, different decade
Seoul National, UIUC, and Largosoft show AI agents misread punctuation in trusted data as structural delimiters. No CVE, no vendor fix planned.
ClickLock macOS stealer kills apps until user types password
Group-IB documents ClickLock, a macOS stealer delivered via ClickFix that kills Finder, Dock, and browsers on a 210ms loop until the victim types their login password.
PhantomEnigma rides Brazilian .gov.br sites and mailboxes
ANY.RUN links a Brazilian banking crimeware operation to 20+ hijacked .gov.br sites and mailboxes, using signature-valid mail and trusted redirects.
AttackerKB's public tier closes August 18
Rapid7 retires the public AttackerKB site and its open submissions on August 18. Analysis, writeups, and API access move behind curation and a customer login.
23andMe settles genetics breach: $18M, 43 states
Multistate AG coalition led by New York's Letitia James. Settlement resolves claims over the 2023 credential-stuffing breach that exposed 6.9M customers' genetic profiles.
Daxin resurfaces in Taiwan alongside new Stupig backdoor
Symantec finds the Daxin kernel rootkit resurfacing at a Taiwan manufacturer, alongside a previously unreported pre-login SYSTEM backdoor called Stupig.
Two Scattered Spider affiliates get 5.5 years for TfL hack
Thalha Jubair, 20, and Owen Flowers, 18, pleaded guilty under the UK Computer Misuse Act. The 2024 intrusion knocked out 148 TfL systems and cost £29 million.
Unpatched Shark vacuums: regional root, no CVE, no patch
tokay0 published a Shark robot vacuum flaw July 13: over-permissive AWS IoT device cert grants root on any other Shark in the same region. No patch.
OpenAI discloses GPT-Red, its internal automated red-teamer
OpenAI describes GPT-Red, an internal automated red-teamer that scales prompt injection discovery and adversarially trains later models against those attacks.
Intruder ships an LLM vuln-discovery product, plus a 0-day
Intruder shipped an LLM code-slicing pipeline that turned up a WordPress plugin zero-day, plus more bugs still under responsible disclosure.
Dutch bust €100M fraud ring, 20 call centers, 700 shills
Dutch Politie takedown of a 2021-active investment-fraud ring — 20 call centers, ~700 fake advisers, five-country arrests, €100M+ estimated peak monthly.
Unit 42: TuxBot v3 shipped LLM chain-of-thought in comments
Palo Alto Unit 42 documents TuxBot v3, an IoT botnet whose developer left an AI safety disclaimer and raw reasoning traces in the shipped binary.
Trend Micro: bandcampro ran a C2 botnet on Gemini CLI
Trend Micro logs 200+ Gemini CLI sessions from a Russian-speaking actor tracked as bandcampro: C2 migration, credential work, and daily botnet ops.
Kaspersky: OkoBot phishes seeds inside Ledger, Trezor apps
Kaspersky's GReAT team says OkoBot has hooked Electron in Ledger and Trezor apps since April 2025 to draw a fake seed-phrase prompt inside the real wallet UI.
Mindgard: Cursor still runs git.exe from repo root
Aaron Portnoy's Mindgard team went public today: Cursor 3.11 on Windows executes any git.exe sitting in a cloned repo's root — seven months, no patch.
Jalisco kit auto-refreshes M365 device codes on demand
ReliaQuest maps two new M365 phishing kits: Jalisco auto-refreshes OAuth device codes to defeat the 15-min window, OmegaLord harvests phones for MFA bypass.
Spain Dismantles €140M BEC Ring; 800 Accounts, 67 Mules
Spanish National Police dismantle a €140M BEC and investment fraud network using 800 bank accounts, 120 companies, and 67 mules; four arrested across three countries.
LastPass, Bitwarden users hit by lookalike-domain phishing
LastPass and Bitwarden users are getting phishing from lookalike "compliance" domains pushing a DocuSign-styled downloader. Delete the email; don't click.
Blackpoint flags LabubaRAT: Rust MaaS RAT poses as NVIDIA
Blackpoint Cyber's Sam Decker and Nevan Beal document LabubaRAT — a Rust MaaS trojan on Windows that ships as nvidia-sysruntime.exe with runtime config.
Grok Build v0.2.93 uploaded whole repos to xAI's bucket
xAI's Grok Build CLI v0.2.93 uploaded whole git repos, history and all, to a GCS bucket. The "Improve the model" toggle didn't stop it. Fix is server-side.
A year of ShinyHunters OAuth abuse, mapped by Microsoft
Microsoft's July 13 report maps three OAuth paths ShinyHunters-linked actors used against Salesforce customers for a year — none of them a Salesforce bug.
Forg365 shows PhaaS became a $400/mo rental market
Analysis: Forg365's $400/mo Microsoft 365 phishing kit adds device code, AitM, and AI-drafted replies. What changed here is finish, not the underlying kind.
NCA charges five over Russian Coms spoofing platform
The NCA charged five London residents over Russian Coms — a caller-ID spoofing platform behind 1.8M scam calls and 170,000 victims. Westminster court date Aug 14.
Meta patent describes an always-on emotion-reading AI
Meta patent 2026/0182881, published July 2, describes an always-on AI that tags voice, biometrics, and app use to score a user's emotional patterns.
Nihon Kotsu cyberattack takes Japan taxi dispatch offline
Japan's largest taxi operator says a July 12 malware intrusion knocked dispatch, web booking, and labor-taxi services offline. No group has claimed.
MemGhost: an email that rewrites an AI agent's memory
arXiv paper: one crafted email talks a memory-enabled AI agent into writing attacker-supplied 'facts' into its memory files. Future sessions load them.
Lidl online shop breach hits DE, BE, NL via provider
Lidl says a file at an unnamed service provider was accessed; DE/BE/NL online shop customer PII taken. Passwords and payment data not yet ruled out.
Huntress Flags Suspected AI-Written PowerShell in AD Case
Huntress attributes an early-June AD enumeration case to a PowerShell script with clear LLM tells — cyan-and-green banners and 'FULLY FIXED' in the title.
First joint EU-UK cyber sanctions name 33 Russian targets
The EU Council named 9 individuals and 4 entities; the UK named 24 more. FSB Center 16, Sandworm, Turla, Lumma Stealer, and Rybar LLC are on the list.
Three Evilginx Crews, One Forgotten Bash History
Lexfo pulled the full toolkit from an open Python server in Budapest and pivoted to two more Evilginx operations targeting Microsoft 365 tenants.
Ill Bloom: Weak PRNG Drained $5.1M From Crypto Wallets
Coinspect's Ill Bloom disclosure: five unnamed wallets shipped seed-phrase code with weak randomness. Two sweeps in May and June drained $5.1M.
China, India APTs Converge on Balochistan Police
SentinelLABS ties 22 months of intrusions at Balochistan Police to two separate crews: China-nexus operators using PlugX and India-linked Mysterious Elephant.
Australia's ACSC names 18 CMS bugs under exploitation
Australia's ACSC named 18 CVEs across WordPress plugins, Craft CMS, Joomla JCE, and more as active exploitation targets, with attackers dropping webshells.
Ghostcommit and the reviewers that don't open the PNG
A PNG carrying prompt injection slips past AI code reviewers that never open image files, then talks a coding agent into exfiltrating a repo's .env secrets as a list of numbers.
Silver Fox ships MODBEACON, a Rust RAT with gRPC C2
QiAnXin attributes a new Rust-based RAT called MODBEACON to Silver Fox, using gRPC streaming for encrypted C2 and SEO-poisoned installers for delivery.
Metasploit Weekly Adds Flowise CSV, macOS PackageKit
Rapid7's Metasploit weekly drops two modules — a Flowise CSV Agent prompt-injection RCE and a macOS PackageKit LPE. New tooling, not new bugs.
Balbooa, iCagenda Join KEV: Four Joomla RCEs in Four Days
CISA added Balbooa Forms and iCagenda to KEV on July 10 — two unauthenticated file-upload RCEs in Joomla extensions. Federal due date is July 13.
A seized crypto account that moved from a cell
Rossen Iossifov, ten years into a laundering sentence, is charged with moving $290K from a seized crypto account. The interesting part is it still moved.
OpenClaw patched a chain that started in a chat message
OpenClaw 2026.6.6 closes three flaws that let a WhatsApp message reach the host as command execution. No public PoC, no observed exploitation.
Politie Points at Dutch Hackers in the 88GB Odido Leak
Dutch National Police say strong indications point at Dutch attackers behind February's Odido breach: a Dutch-speaking vishing call to customer service, then 6.2M records leaked.
A laser resets Tangem wallets, and there's no patch
Ledger Donjon's laser fault-injection attack resets a Tangem card's password without the old one. There is no patch — Tangem ships no firmware updates.
XRING: 260 bytes, no patch, three months of Alibaba silence
FoxIO's Sébastien Féry disclosed a QPACK integer underflow in Alibaba XQUIC that crashes HTTP/3 servers with 260 bytes. Reported April 7. No reply. No patch.
WP-SHELLSTORM ran 22 days with its door left open
SOCRadar and Ctrl-Alt-Intel pulled 22 days of files off an exposed WP-SHELLSTORM server: 1.4M targets, 25K compromises, 5,700 live shells.
Ill Bloom is a $3.1M lesson in weak randomness, again
Coinspect disclosed weak PRNG in wallet recovery-phrase generation; attackers drained $3.1M in a May sweep. The pattern — bad randomness, stolen keys — is old.
Meta's Muse Image defaults on for public Instagram
Meta's new Muse Image model reuses public Instagram photos and reels by default — no notification, no watermark discussion, opt-out three levels deep in Sharing settings.
The clearinghouse boom is not new, and neither is the fatigue
Chainguard announced Athena. Red Hat and the White House announced Lightwell. Vulnerability clearinghouses have been getting reannounced since the 1980s.
The ATO fight moved past credential stuffing
The Hacker News argues account takeover shifted from credential stuffing to attacking verification — passkeys pushed the front door shut, so attackers moved.
Talos on 'attackers only need to be right once'
Cisco Talos's Hazel argues 'attackers only need to be right once' is a cliché the defensive community should retire. It's overdue.
Datadog: 50+ dormant GitHub accounts mapping org charts
Datadog Security Labs documents 50+ dormant GitHub accounts running months-long enumeration of corporate orgs, repos, and — in some cases — private code.
Helix: new data-extortion crew hits SharePoint via vishing
ReliaQuest attributes new data-extortion crew Helix to vishing and device-code phishing against SharePoint. Infrastructure overlaps BlackFile.
GigaWiper/BLUERABBIT: Go-based wiper, CyberAv3ngers-linked
Microsoft and Binary Defense concurrently disclose a Go-based Windows destructive backdoor — wipe, fake ransomware, spyware in one binary — attributed to Iran-nexus CyberAv3ngers.
INTERPOL First Light 2026: 5,811 arrests, $293M seized
INTERPOL's Operation First Light 2026 arrested 5,811 fraud suspects across 97 countries, seized $293M and blocked 31,014 accounts over 3.5 months.
Friendly Fire: agents review the trap, then execute it
AI Now Institute researchers show autonomous Claude Code and Codex can be tricked into running a hidden binary during their own security-review pass.
AssuranceAmerica breach: 6.9M drivers, 4-month notice gap
AssuranceAmerica confirms a March 16 intrusion exposed data on 6,998,886 drivers. Notification letters went out in July — a nearly four-month gap between detection and public notice.
Infoblox: Lurking Lizard runs 230-domain fake 7-Zip proxy
Infoblox ties a China-based residential-proxy operator to 230+ lookalike domains active since 2022, seeding fake 7-Zip and WireVPN installers.
GhostApproval symlink bug hits six AI coding assistants
Wiz research: Amazon Q, Cursor, Claude Code, Augment, Antigravity, Windsurf all approved one file path in the dialog while writing to another via symlinks.
Spain arrests suspected CARR logistics operator
Spanish police detained a Palencia man tied to CyberArmy of Russia Reborn, Z-Pentest, and NoName057(16). The announcement lands nearly four months after the raid.
Krebs traces zero-day broker IRIS C2 to Wohl and Burkman
Krebs ties IRIS C2, an offensive-security startup pitching zero-day acquisition, to Jacob Wohl and Jack Burkman — both convicted of felony fraud.
Sophos: Coding Agents Are Tripping the Attacker Detections
Seven days of Sophos endpoint telemetry: Claude Code, Cursor, and Codex trip the same rules built to catch attackers — because behaviorally, they should.
Refused in Chat, Written in Code: Copilot's Workflow Gap
Kumar and Maple's new arXiv preprint says Copilot's Claude and Gemini backends refused harmful prompts in chat but produced them 816-for-816 in a workflow.
SCMBANKER active against Mexican banks — Elastic REF6045
Elastic Security Labs is tracking SCMBANKER (REF6045), a PowerShell fraud toolkit hitting Mexican banks, fintechs, and crypto exchanges via ClickFix lures.
KDDI Breach: 12M Emails, 7.6M Passwords via 3rd-Party 0day
KDDI says a May 16 zero-day in unnamed third-party software exposed 12,233,087 email addresses and 7,616,173 passwords across five Japanese ISPs.
CISA Adds Langflow and Two Joomla Builders to KEV
CISA added three vulnerabilities to KEV on July 7 — a Langflow IDOR and two Joomla page-builder RCEs. Federal due date is July 10. Priority order below.
Proofpoint: China cluster raids university physics mail
Proofpoint attributes a Roundcube-exploitation campaign against U.S. and Canadian university physics departments to a China-aligned cluster, UNK_MassTraction.
China-Linked UAT-7810 Expands ORB Net With LONGLEASH
Cisco Talos ties China-aligned UAT-7810 to LONGLEASH backdoor and an expanding ORB relay network built on unpatched Ruckus and ASUS routers.
Windows Device ID trail led FBI to Scattered Spider suspect
A newly unsealed federal complaint says a Microsoft-recorded device ID tied the account behind a Scattered Spider intrusion to 19-year-old Peter Stokes.
Accenture Confirms Breach; Attacker Claims 35 GB Stolen
Accenture confirmed a security incident. A threat actor is advertising 35 GB of alleged source code for sale. The volume claim is unverified — treat accordingly.
DragonReturn Drops DcRAT on Indian Taxpayers
Seqrite Labs attributes an ongoing spear-phishing campaign against Indian tax filers to a suspected China-nexus actor with infrastructure and tactical overlap to Silver Fox. First observed May 18.
QuimaRAT: A $150 Cross-Platform Java RAT MaaS
LevelBlue profiled a new cross-platform Java RAT sold as MaaS. No confirmed campaigns yet — but the price is low, the payload runs everywhere, and the loader is built to walk past SmartScreen. Assume it lands somewhere soon.
Flipper Zero Firmware Goes Maintenance-Only
Flipper Devices says the Flipper Zero firmware is stable at 1.0 and full-time feature work is over. Community PRs run the future, filtered through GitHub Discussions voting and stricter review. Here's what changes.
Metasploit's July 3 Drop: SMB-to-Meterpreter, Peyara
Rapid7 shipped an SMB-to-Meterpreter session upgrade and a Peyara Remote Mouse RCE module this week. Neither is novel research. Both change what your alerts will look like. Here's the tune.
IGA Was Built Around Employment Records, Not Agents
A contributed piece to The Hacker News from Orchid Security lays out where the joiner-mover-leaver model quietly fails for AI agents. Vendor-adjacent, but the gap analysis holds.
Talos on Curiosity: A Skill That Doesn't Scale
William Largent's Threat Source column this week reads as an essay on board games and pattern recognition. It's really an argument about the load-bearing skill that keeps a defender from becoming a checklist.
PamStealer: A Fake Maccy Site Steals macOS Creds
Jamf Threat Labs disclosed a new macOS credential stealer today that impersonates the Maccy clipboard app, validates the victim's login password against PAM in real time, and exfiltrates keychain and browser data. Apple Silicon only. Here's what defenders should do.
FBI Seizes NetNut Proxy, Google Degrades Popa Botnet
The FBI seized hundreds of NetNut proxy domains on July 2; Google's Threat Intelligence Group, working with FBI and Lumen, cut the linked Popa botnet's usable device pool by millions the same day.