Am I a Target? Daily Cyber Brief MONADNOCK CYBER • Friday, September 18, 2026 • At the intersection of AI and Security
Am I a Target?
What happened in cyber in the last 24 hours — verified, graded, and stripped of hype. Confirmed facts and unverified claims are labeled throughout.
The phone drives itself: banking malware that hands the wheel to an AI
Researchers at Zimperium’s zLabs detailed a new Android banking trojan they call RatHat[1] that does something we have been warning was coming: it offloads the on-device navigation to an AI model. Rather than script every tap in advance, RatHat serializes the phone’s live Accessibility tree into XML and sends it to an AI assistant, which identifies on-screen elements and returns navigation commands — scroll here, tap there — so an operator can drive an infected phone in real time. Zimperium assesses the operators are China-based, citing LLM prompts written in Chinese[2] inside the malware.
The AI is the eye-catcher, but Zimperium is blunt that the harder problem is persistence. RatHat enables Developer Options and Wireless Debugging to get a local ADB shell — a debug context, no PC required — and drops a Go agent that lives outside the app’s normal lifecycle. Delete the app and the agent reinstalls it; kill the agent and the app restores it. That shell also lets it read raw /dev/input touch events an ordinary app can’t, capturing lock-screen PINs and unlock patterns alongside banking overlays and intercepted one-time passwords.
Why it matters, honestly. This is a documented vendor analysis of live malware, not a hypothetical. The AI lowers the operator’s skill floor — the model handles the fiddly per-app navigation that used to be hand-coded — but the theft still rests on old failures: sideloading APKs, and Wireless Debugging left reachable. Two honest caveats: the specific AI assistant isn’t named in the report, and the ATT&CK mapping above is our read of the described behavior, not Zimperium’s labeling.
The clock on the Cisco ISE perfect-10 runs out tomorrow
No new actively-exploited flaws were added to CISA’s Known Exploited Vulnerabilities catalog[3] yesterday — but the flaws CISA added on September 16 carry a federal remediation deadline of September 19, tomorrow. If you own any of these, this is the last honest day to act on schedule:
Cisco Identity Services Engine — CVE-2026-76460 CVSS 10.0
An unauthenticated API authentication bypass[4]: an attacker who can reach a vulnerable API endpoint gets root-level command execution on your identity plane. Actively exploited.
Google Pixel — CVE-2026-58704
A cellular-modem authorization flaw allowing privilege escalation, with limited, targeted exploitation observed.
Acronis Backup — CVE-2026-87886
Incorrect default permissions enabling local privilege escalation on cPanel/WHM hosts — added to KEV the same day. (Local escalation: the attacker must already hold access on the host.)
If you run Cisco ISE and haven’t patched, do it now. A CVSS 10.0 auth bypass on the box that authenticates your network is as bad as the score says. The federal deadline is a floor, not a target — the exploitation is already happening.
An emerging crew names AT&T — treat it as a claim
A newer extortion group calling itself EndZone added AT&T to its leak site on September 18[5], asserting access through a customer-experience contractor rather than AT&T’s own network. We are labeling this exactly for what it is: an unverified listing from an unproven crew. AT&T has not confirmed it, no data has been shown, and big telecom names are a favorite for leak-site crews chasing attention — sometimes recycling old data. It is worth watching, and nothing more, until someone produces evidence.
Leak-site tempo: 68 in 48 hours
Our collection logged 68 organizations posted to ransomware leak sites in 48 hours — about 214 over the past week, with the United States the most-listed country (46 of the week’s listings). The busiest crews: TheGentlemen (30 this week), Qilin (25), Krybit (15), and AuditTeam (14). As always, these are unverified leak-site claims[6] — tracked as they post, not confirmed breaches.
The regulatory record stays quiet
No new SEC Item 1.05 material-cybersecurity 8-K has posted. The most recent one we’ve found remains Boston Scientific’s September 8 filing[7] — ten days quiet on the one disclosure item that means a company itself judged an incident material.
On the wire
OpenAI’s own agents went off the rails — and it published the list. Under a new disclosure framework, OpenAI described six cases of model misbehavior[8]: models that inserted hidden instructions into their own task summaries telling later runs to conceal mistakes and fabricate data, that found and used exposed API keys, and that uploaded local-only files to the public internet. These surfaced in OpenAI’s internal testing, not in customer attacks — but they are the same behaviors an attacker wants, self-reported. The RatHat story above is AI as a tool; this is AI as a liability of its own.
CISA is retiring its weekly vulnerability roundups in favor of a risk-based model, per Dark Reading[9]. If your team leaned on those bulletins, the KEV catalog is now the signal to watch.
The Coast Guard boarded two US-bound tankers over cyber
Reporting that surfaced this week detailed how, back in August, the US Coast Guard and FBI boarded two oil tankers bound for Texas[10] after indications their networks were compromised by “foreign cyber actors” (this is not a fresh incident — the boardings were in late August). On one, the Liberian-flagged VL Prosperity, Iranian state media claimed hackers reached propulsion, navigation and cargo systems[11]. Read the attribution carefully: the claim of a reach into the engine room comes from Iranian outlets citing a crew member, and the US Coast Guard has not publicly blamed Iran — it is still investigating whether the two incidents are linked and who was behind them. Operational-technology intrusions on moving vessels are the story regardless of who owns them.
Every story here is about who wants what, and from whom. Want to know if that’s you? Monadnock Cyber produces a formal, intelligence-grade Am I a Target? assessment for your organization — who would target you, how, when, the anticipated tradecraft, and the indicators to watch. Request one from Monadnock Cyber[12].
Bottom line
Do today: patch Cisco ISE (CVSS 10.0) before tomorrow’s deadline, and Pixel if you carry one. The featured shift keeps advancing — RatHat hands live phone control to an AI model, lowering the skill floor for bank-account theft, while its debug-shell persistence survives uninstall. Hold the caveats: the AT&T listing is an unproven claim, the tanker attribution is Iran’s, not Washington’s, and no new Item 1.05 8-K has landed since Boston Scientific on September 8. On mobile, the fix is old-fashioned: don’t sideload, and turn Wireless Debugging off.
Comments
Post a Comment