Leak exposes undercover ICE agents

Hi there, this is your daily ☕️ Cyberpresso.

Leak exposes 41,000 Homeland Security staff LINK

• A leaked archive called Icepick has exposed records for more than 41,000 Homeland Security staff and contractors, revealing contact details, job assignments, and links to sensitive work, including undercover ICE operations.
• Released on October 28 by Distributed Denial of Secrets, the files cover DHS and seven agencies, with CBP making up over half and about 10,000 records tied to ICE, roughly one seventh of the department's workforce.
• WIRED found 39 Homeland Security Investigations staff in permission groups marked "CUC," a term court records link to certified undercover work; DDoSecrets blames an Okta sign-on service, but Okta says it found no evidence of any compromise.

Cisco flaw lets attackers reboot firewalls LINK

• Cisco patched a firewall bug, tracked as CVE-2026-76418, that attackers are already abusing to force Secure Firewall ASA and FTD devices to restart, knocking out VPN access without any valid login.
• The high-severity flaw, scored 8.6, comes from logic errors in the Remote Access SSL VPN service when secondary authentication is set on a tunnel group; a crafted request triggers the crash but causes no code execution or data theft.
• Cisco posted its advisory yesterday with fixed ASA and FTD releases and no full workaround, only temporary steps around tunnel group settings, urging admins to check for webvpn and secondary authentication and patch quickly.

China hackers exploited VMware flaw LINK

• A China-linked hacking group called Warp Panda exploited a flaw in Broadcom VMware ESX to hide its tracks while attacking U.S. law firms, prompting CISA to add the bug to its known-exploited list.
• The bug, CVE-2026-41709, is an insufficient-logging flaw that lets an attacker with admin access run operations without being recorded; Warp Panda used it as a zero-day in early 2026, and Broadcom shipped fixes in July 2026.
• After breaking in through internet-facing edge devices, Warp Panda pivots to VMware vCenter to plant malware like BRICKSTORM, steals cloud session tokens, and registers rogue MFA devices; federal agencies were told to patch by November 11, 2026.

Edge exploit escapes sandbox LINK

• Security researcher Orange Tsai revealed a zero-click attack on Microsoft Edge that strings together five bugs and two Windows features to run code outside the browser's protective sandbox, turning a malicious webpage into Windows code execution.
• Built for Pwn2Own Berlin 2026 and published yesterday, the chain begins with a flaw in Edge's Conditional Access navigation that creates universal cross-site scripting, then abuses the Immersive Reader and a config file to escape.
• Tsai's exploit used fewer than 100 lines of code and no memory corruption; there is a public write-up but no reported real-world attacks, and Microsoft's advice is to keep Edge updated through its security releases.

Strategies & Tactics

> How attackers keep reading mail after M365 session revocation: Revoking sessions or resetting passwords on a phished Microsoft 365 account leaves stolen tokens working for mail-reading APIs, so defenders must disable the account entirely.

> Cloudflare Image Transformation Flaw Allows Hackers to Bypass CSP and Execute Malicious JavaScript: Researchers chained an SVG sanitizer bug with cached-header manipulation to run malicious JavaScript on Cloudflare-served sites, bypassing the browser's content security policy.

> Attackers turn GitHub issues into supply chain campaign against 61 public repos in less than an hour: Attackers hid malicious commands in GitHub issue text to trigger automated build pipelines and steal credentials, so configure workflows to distrust external content.

> Managed for Malice: How Attackers Turn MDM into Remote Access: Attackers trick users into enrolling devices into mobile management tools, gaining a trusted path to reinstall remote-access software, so treat unexpected enrollments as security events.

> Pluto Security details how a trust_remote_code download turns into malware: Setting trust_remote_code=True runs a model's bundled Python at load time, so attackers hide miners and persistent backdoors in code scanners never inspect.

Other news you might like

• Attack on email hosting provider for German far right party AfD exposed 345k accounts LINK

• Hacker Drains Nearly $1,000,000 in SOL From Solana Trading Bot Users in 94 Seconds LINK

• GitHub Launches npm Firewall to Block Malicious Packages Before Installation LINK

• Imposter abused trust in OpenAI, Anthropic, and 30 others to hunt for secrets LINK

• Chrome extensions with 13.4 million users funnel searches through tracking servers LINK

• Officials seize domain of top bulletproof hosting provider Virtualine LINK

• Researchers uncover Volt Typhoon’s covert, ephemeral network LINK

• Unknown APT uses NetamLoader to attack a Win10 OS customized for governments in Asia LINK

Trending tools

Harden: a free, local security tool that checks AI coding agent tool calls before they run, keeping your repo and output on your machine. LINK

CtrlOps: a local-first desktop app for managing Linux servers with an AI-assisted terminal, real-time monitoring, and agentless SSH access LINK

VibeDefend: embeds security and business rules into AI coding agents like Cursor and Copilot, catching vulnerabilities and malicious packages before code reaches your PR. LINK

Axari: an AI security twin that works across your tools and teams, taking goals and recurring tasks to completion directly within Slack or Teams. LINK

Upsolve Row-Level Security: enforces row-level data permissions server side so AI agents can serve every tenant without ever seeing or bypassing access rules. LINK

AuthMonster: a two-factor authentication app with biometric locking, simple migration from Google Authenticator and LastPass, and a donation-based model instead of paywalls. LINK

Trending papers & reports

Cloud storage audits let data owners verify a provider has not altered or deleted files by checking each one against a single tamper-proof fingerprint on Ethereum, catching every change tested with no false alarms. LINK

Reverse-engineering assistants that must run on small, locked-down machines got a training boost, lifting a compact open model to ~95% accuracy at predicting how code behaves, nearly matching frontier models at a fraction of the size. LINK

Hidden-instruction theft can be caught by spotting the pattern across a series of probing questions, catching nearly all attacks that trick chatbots into revealing their secret setup instructions. LINK

Prompt injection defenses teach an AI model to read untrusted text without obeying hidden commands in it, cutting a top attack's success rate from 97% to 10% while preserving normal performance. LINK

See you tomorrow for a new dose of ☕️ Cyberpresso!