Pinned Tweet
Hundreds of cybersecurity blogs, research reports, and advisories published every day. No one has time to read them all. And the one report that matters? It's buried somewhere in the noise.
That's why DFIR Radar exists.
We monitor the cybersecurity landscape around the clock. Every article is evaluated for DFIR relevance. Only what's genuinely useful makes it through. The rest never reaches your feed.
This feed is the result of that process. Every article is sourced, evaluated, and published only if it meets the standard.
If you find something we missed, our Discord community lets you contribute directly.
Discord community: discord.gg/rHkqgs53bF
Built by a practitioner who needed this to exist.
Follow once. Stay informed forever.
#DFIR_Radar
ARTEX, an open-source AI-agent pen-testing framework, became an attacker staple in about three months. Storm-0940, assessed by Microsoft as China 🇨🇳 state aligned, used it to hit a vendor's Microsoft 365 apps, then tried sign-ins against more than 2,250 customer tenants.
- Storm-0940 chain (Microsoft's Oct 21 report): a password spray from 144.172.64[.]99 ran Oct 10 to 14, with hundreds of thousands of failed attempts peaking at thousands per minute on Oct 11. After taking over the vendor's app identities, the actor signed in as service principals and ran Microsoft Graph recon. It added credentials to existing OAuth apps, read secrets across multiple Key Vaults, and in one tenant used a management app to assign itself Global Administrator. No Microsoft vulnerability was involved, only valid credentials. Attribution is Microsoft's assessment. Google tracks overlapping activity as UNC7228, dating back to at least August 2026, hitting US 🇺🇸 and European entities in enterprise software, industrial automation, semiconductors and telecom. Wiz reports the actor now targets IT and software vendors, focusing on identity providers such as Okta and Entra and on Microsoft 365 integrations.
- Machine speed shapes the evidence. Microsoft saw one agent go from searching GitHub to exfiltrating Azure Key Vault data in five minutes. Google saw bursts of 2 to 25 seconds, including 27 events in 23 seconds. Failed calls get fixed fast: a wrong Key Vault API version was retried correctly about 28 seconds later, Azure SQL firewall rules after about 60 seconds, Application Insights after about 80. Agents rotate fingerprints once blocked, and Wiz saw more than 15 user-agent strings from Storm-0940-linked activity.
- Storm-3358 shows the non-cloud side: anonymous file upload and SSRF flaws on internet-facing systems, then agents pulled WISP, vshell and AdaptixC2 with netcat and curl. Reported ARTEX servers: 91.92.240[.]238, 85.11.168[.]45, 178.16.54[.]129, 31.77.229[.]35, with C2 at 85.11.167[.]52 and WISP nodes at 199.19.95[.]66, 199.19.95[.]67 and 199.19.95[.]68.
- Tool-level artifacts: files named artex-probe-* with roughly incrementing five-digit numbers, and created accounts with ArtexProbe in the name and plaintext passwords. Short-lived JSP webshells include cl1217.jsp, f1219a.jsp and z1219.jsp; Google saw one written through MySQL general log poisoning. C2 includes rotating webhook[.]site URLs, 491 of them in one intrusion Google observed. Client strings include curl/7.88.1, python-requests/2.34.2, Python-urllib/3.12, norma/0.4 and ReconProbe/1.0, though Wiz cautions many are not unique to ARTEX. Cloud behavior to look for: one source authenticating as multiple service principals, or one app requesting tokens for several Azure services in a short window.
- Entry often starts with leaked secrets. Apiiro found 2,539 sensitive secrets across 1,994 public GitHub repos in AI coding agent settings files, covering database passwords, API tokens and root logins. Microsoft says agents repeatedly focus on credentials that reach AI services. With more than 12,000 forks, shared defaults will not tell you which group is behind an intrusion, and Google forecasts fully autonomous, stealthier attacks within 3 to 6 months.
Takeaway: these sign-ins do not show up in user sign-in logs. Collect service-principal and non-interactive sign-in logs for every tenant and hunt weeks back. Inventory third-party multi-tenant apps across tenants, check for credentials added to existing OAuth apps, and alert on any app assigning itself Global Administrator. Treat any publicly exposed credential as compromised. Network IOCs are short-lived, so build detections on behavior: a failed secret followed by a success, one app requesting many token types in minutes, retries corrected within a minute, and dozens of events in seconds. The IOC lists from Microsoft, Google and Wiz are in the report.
#DFIR_Radar
1
1
85
Bazuka, found during an NFC relay investigation, is a full Android fraud framework: dropper, core, 23 modules, ADB bridging, and LLM agents in the operator panel that help pick victims.
- Delivery: Traffers push the dropper through landing pages with lures like IPTV apps and adult content. A VirusTotal sample kept its filename, TikTok18-Dropper-tiktok18-premium-1793332000-daf6efe2.apk, SHA256 f046b50166c8229b49839eeebcc7fd14bf3f198e4ec1bfde8a5b271e1407dec5. It asks for install-from-unknown-sources permission, installs a stager through a PackageManager session, and launches com.tiktok.premium[.]pro. A second variant is a downloader, SHA256 4d572669d6000304a692874c1583021dfdb4f2f113ea487e470f7e4c486e503a, which pulls the core directly. The dropper builder uses minimal manifest permissions, apparently to help pass Google Play review. ThreatFabric has seen no sign that this happened.
- Loading and C2: The core and modules are DEX files loaded by InMemoryDexClassLoader straight from a byte array, so modules never touch disk. The stager calls DualWebSocketManager, WireSessionManager and ModuleLoader under com[.]support.assistant[.]network. The core talks over Control and Data WebSockets with ChaCha20-Poly1305. Keys rotate, sit in SharedPreferences, and are also sent to C2. On connect, the bot reports device info including a ThreatLevel, so operators can rate a victim within seconds.
- Permission abuse: The core taps through prompts itself, using Accessibility or ADB shell commands. It can enable Developer options and Wireless Debugging, intercept the pairing PIN, and open a TLS session to the ADB daemon. Accessibility rebinding paths include the enabled_accessibility_services secure setting, the self_adb module's rebind_accessibility command, and the com.guidocool.cryptomanager.action.A11Y_REBIND broadcast. The uninstall_shield module counters removal, and self-destruct calls setApplicationEnabledSetting(packageName, 3, 1). A new overlay trick screenshots the home or lock screen and replays it on command.
- Fraud modules: input_recorder is a regex keylogger for card numbers, wallet addresses and passwords. twofa and smsmessage grab one-time codes. clip_monitor swaps BTC, ETH, TRC20 and SOL addresses. screen_control captures the lock PIN, password or pattern. optics handles remote control and phishing overlays, proxy is a reverse TCP proxy, and tapread-nfc_module relays contactless cards. The report says automated payments are the one attack the framework does not cover.
- Infrastructure: Cloudflare-fronted subdomains of brokerok[.]org are hardcoded in the code and in the downloader's bundled certificate: c2ws.brokerok[.]org, c2data.brokerok[.]org, c2proxy.brokerok[.]org, c2api.brokerok[.]org. Two origin servers turned up. 95.181.213[.]21 sat in Moscow and was already offline. 170.168.61[.]117 (gargona[.]cash) sits in Amsterdam, Netherlands 🇳🇱, hosted by GTHost, and its address is in the dropper's certificate. Ports: 8082 and 8445 for the panel, 8443 for bot C2, 8444 for the data channel. The panel frontend is readable without authentication, mixes in Russian 🇷🇺 UI strings, and embeds a console plus LLM chat that can score victims and suggest permission-click strategies.
Takeaway: Hash hunt with the core, dropper, downloader and per-module SHA256 values in the report appendix. Check devices for com.tiktok.premium[.]pro, unexpected Accessibility service entries, Wireless Debugging switched on without user intent, and traffic to the brokerok[.]org subdomains or the two origin IPs. Because modules load only in memory, device triage should favor package, settings and network evidence over file carving.
#DFIR_Radar
1
2
149
Microsoft's agent-framework on PyPI (about 40,000 weekly downloads) was hijacked on November 17, 2026. Versions 1.26.1 and 1.26.2 ship a .pth startup hook that runs a credential stealer on every Python launch, even if your code never imports agent_framework.
- Timeline and provenance: 1.26.1 went live at 20:35 UTC and was quarantined at 21:06. The attacker uploaded 1.26.2 at 21:08, and both were gone by 21:16. Renovate was opening upgrade PRs within six minutes of 1.26.1. This was not a Microsoft pipeline compromise. ZIP headers show Info-ZIP zip 3.0 repacking at 20:26:19 UTC under uid 1000 on a UTC+8 host, while the WHEEL file still claims flit 4.2.0. JFrog has not attributed it to a known group.
- Why the hook works: the agent-framework meta-package only installs a stub, agent_framework_meta, and nobody imports it. So the attacker left the agent-framework-core[all]==1.26.0 pin intact and added a 28-byte agent_framework_meta.pth to the root of site-packages containing "import agent_framework_meta". site[.]py exec()s that line at interpreter startup, so python, pytest, IDE language servers, container entrypoints and even pip trigger it. The 1.26.1 hook calls _telemetry._start(). In 1.26.2 the clean __init__.py was restored and the hook moved to af_runtime_diag.pth and af_runtime_diag.py. It forks and calls setsid on Linux and macOS, and uses a daemon thread on Windows. _AF_TEL=1 prevents fork loops, and no lockfile is written, so every fresh launch exfiltrates again.
- What it steals: environment variables matching 31 substrings (Azure, AWS, GitHub, OpenAI, Hugging Face and more), plus 35 file globs. These cover ~/.azure/*.json token caches, .npmrc, .netrc, .git-credentials, .docker/config.json, /app/.env, /workspace/.env and Azure DevOps agent .credentials stores. On Linux it reads the first 300 /proc/[0-9]*/environ entries. It also queries IDENTITY_ENDPOINT, MSI_ENDPOINT or the VM metadata service at 169.254.169[.]254 for live OAuth2 tokens across 12 audiences. 1.26.2 added GitHub Actions OIDC tokens and /home/runner paths.
- Exfiltration and operator mistakes: data is JSON, zlib, Base64 with no encryption, sent by HTTPS POST to hxxps://webhook[.]site/d381fd80-ec62-42ad-8f5c-d1b15103c9bb?t=telemetry. The User-Agent is af-runtime-diagnostics/1, changed to runtime-diagnostics/1 in 1.26.2. There is no second stage, no persistence outside site-packages and no worm. The inbox is now rate-limited (HTTP 429), but early victims got HTTP 200 and their secrets may be readable by anyone holding the UUID. The Azure DevOps OIDC request is broken, though SYSTEM_ACCESSTOKEN still leaks via os.environ. The IMDS timeouts delay exfil about 104 seconds on non-Azure Linux. Windows CLI runs can exit before the thread finishes, and macOS misses /Users/* paths but still leaks env vars and ~/.azure/*.json.
- Cleanup trap: do not run bare pip uninstall first, because pip is a Python entrypoint and triggers the stealer once more. Export _AF_TEL=1, delete the four malicious files from every site-packages directory using the environment's own Python binary, then reinstall agent-framework==1.26.0. uv does not execute .pth hooks during removal. Rotate everything reachable. Rotating passwords alone does not revoke cached Entra ID refresh tokens, and managed identity or OIDC access tokens stay valid up to 24 hours. Also note that __pycache__ .pyc files do not show whether the payload ran, since pip compiles bytecode at install time.
Takeaway: hunt for agent_framework_meta.pth, af_runtime_diag.pth and af_runtime_diag.py in site-packages, including inside container images built on November 17 and in lockfiles or Renovate/Dependabot PRs from that day. Check build-host DNS for webhook[.]site and TLS-decrypted proxy logs for POSTs to /d381fd80-ec62-42ad-8f5c-d1b15103c9bb?t=telemetry, where HTTP 200 means exfil completed. Review Entra ID Managed Identity sign-ins for token replay. Wheel and file SHA-256 hashes are in the report's IOC tables, so search internal mirrors for the wheel hashes and evict hits.
#DFIR_Radar
1
1
117
AI agents weaponized PHP2Shell (CVE-2026-109618, CVE-2026-109619) the day public PoCs dropped on October 21, 2026. SOCRadar found the operator's exposed C2, tooling and loot: 595 IPs attacked in 57 countries, 404 confirmed compromised.
- Exploit chain: A multipart POST of about 8 MiB with over 1,000 fields exhausts the default 128M memory_limit. A second request on the same keep-alive connection uses dangling session state to leak addresses and call a forged HashTable destructor mapped to zif_system. It works against default PHP/Apache under mod_php or FrankenPHP, not FPM. Fixed releases stop it (8.4.27 cut the stage-1 leak from 4,096 to 32 bytes). Look for HTTP 500 with X-Powered-By: PHP, a second POST on the same connection, and a Set-Cookie PHPSESSID of about 4,096 or exactly 16 bytes. On Nextcloud, POSTs to /status.php, /cron.php and /public.php.
- AI layer: A self-built queue (maq) ran 39 parallel omp workers against two vLLM DeepSeek V4.1 endpoints. A GOAL[.]md told an agent to work autonomously toward over $50,000 USDT. SOCRadar assesses the operation as human-directed and AI-augmented, and full autonomy was not confirmed. No crypto theft was confirmed. The completed impact was about 3.3M user records pulled from an unauthenticated Redis. An OPSEC failure also exposed the operator: a worker ran collect[.]sh on the operator's own lab, and the snapshot was uploaded under job 176 to the public loot service.
- C2 footprint: 206.245.153[.]56 runs an unauthenticated Python file service on 8088. GET /tools/ serves collect[.]sh, deep-loot[.]sh, persist2[.]sh and php-loot.php, and POST /<job>/<name>.tgz returns 201. Port 80 serves /j<job>.sh for curl | sh. Each victim gets a reverse-shell listener on port 41000 + job ID. Related hosts are 157.180.96[.]7 (the AI lab, open directory on 18082) and 185.65.134[.]250 (exploit and SSH source). The c2-agents SSH key is reused across jobs.
- Persistence to hunt: /tmp/beacon.sh (60-second reconnect loop), per-minute cron lines (T1053.003), and added authorized_keys entries (T1098.004). Also look for token-protected PHP webshells named doc/thumb-<8 hex>.php (T1505.003), a Nextcloud Beacon app, a Laravel route-cache backdoor, and /tmp/.c<id> lock markers. A localhost-only Dropbear on 127.0.0.1:2222 is kept alive via a rogue auto_prepend_file, with files under /var/www/html/.p/.
- Impact: The collector bundles secrets into a fixed layout (meta.tsv, inventories from 00-identity to 23-crypto-candidates) and POSTs it, logging UPLOAD=201. Loot includes ECS/IMDS role credentials (T1552.005), .env files, database dumps, git history and session files, about 290,000 files and 9.5 GB in total. The operator documented shells on 217 hosts but root on only 12, mostly unprivileged www-data in containers. Victims were chosen by PHP/Apache fingerprint, not sector. SOCRadar's read is an initial-access-broker pattern, with low confidence on identity.
Takeaway: Patching closes the entry point but not the foothold. Upgrade to PHP 8.2.35, 8.3.36, 8.4.27 or 8.5.12 and verify the SAPI version, not just CLI php -v. Then search web-server hosts for the paths and markers above. Hunt for web-server children spawning sh, curl or bash -i, collector bundles left in /tmp, and outbound traffic to 206.245.153[.]56:8088 or ports 41000-49153. In cloud logs, look for web-app roles calling STS, S3, Secrets Manager, SSM or RDS from non-VPC IPs, and metadata requests from web processes. Rotate every secret on exposed hosts and audit cloud logs back to October 21, 2026. The report's IOC section lists hashes, SSH key fingerprints, paths and URL patterns, plus a detection and hunting section.
#DFIR_Radar
2
1
192
Source: socradar.io/blog/php2shell-a…
41
Attackers abused Optimizely Web to push malicious JavaScript through legitimate sites of major brands, hitting 178 organizations per Report URI. Detect FYI turns the IOCs into KQL hunts for Defender XDR and Sentinel.
- Report URI published the research on November 16, 2026. The injected JavaScript was tied to cryptocurrency address replacement, payment card skimming, information theft, and delivery of the iOS exploits Coruna and DarkSword. Affected brands include BBC, CNN, Zoom, GitLab, IKEA, and Ledger. Users could be exposed just by browsing a trusted site.
- Infrastructure to hunt in network telemetry: 1688360[.]cc, 5vobauua3ojpklx[.]icu, eil786npto9u5bq[.]icu, and webshell[.]lol. Detect FYI's labels for them: webshell[.]lol is JavaScript injection infrastructure, and the two .icu domains are iOS exploit routers. On 1688360[.]cc, track.js is the JavaScript payload, hit.php suggests cookie and screenshot exfiltration, and card.php suggests card data exfiltration. These are the hunter's interpretations, not confirmed behavior.
- Hunt flow in DeviceNetworkEvents: match RemoteUrl against the four domains, then summarize per device to see connection counts, first and last seen, and which activity types each device touched. Repeated hits suggest continued exposure, not confirmed compromise. Pivot on InitiatingProcessId plus InitiatingProcessCreationTime so you track the exact browser instance, not just the process name.
- Follow-on pivots: browser-spawned processes within 10 minutes of a hit in DeviceProcessEvents, and other connections from the same process within 5 minutes. A file-creation join in DeviceFileEvents (2 minute window) produced too many false positives for the author. Expect normal browser background traffic in the network pivot, and note that JavaScript can run entirely inside the browser without spawning anything.
- Infection-path query, added November 17: for the window November 16 to 18, it joins visits to the affected brand domains to malicious-domain hits on the same device in the two minutes before the hit. That points to which trusted site delivered the script. It checks only 1688360[.]cc and the two .icu domains, so add webshell[.]lol if you want the injection host covered too.
- Key limitation: DeviceNetworkEvents does not capture full HTTP/HTTPS payloads. A connection to this infrastructure cannot prove the JavaScript executed or that data left the device. Proxy logs, browser artifacts, or packet captures with decryption are needed to close that gap.
Takeaway: treat the queries as starting points, not production detections. Run the domain match over your retention window first, use the pre-hit brand-visit join to find the delivery site, then pull proxy and browser evidence for any device that hits card.php or hit.php paths. The full KQL set is in the Detect FYI post, and the original IOCs are in Report URI's research.
#DFIR_Radar
1
1
223
Attackers abused CyberPower PowerPanel Business, a UPS management tool, to run SYSTEM commands on a domain controller and collect ntds.dit. Every command traced back to ppbed.exe, not a new service or scheduled task.
- Why it works: PPB can run a .cmd file from its extcmd folder when a power event fires. The attacker used the web console (embedded Tomcat, TCP 3052) to push their own scripts, and the PowerPanel service, ppbed.exe, ran them as SYSTEM. No memory-corruption exploit was needed. In process telemetry it looks like a power product launching cmd.exe, which is easy to dismiss.
- Entry point: requests hit /local/rest/v1/file/upload on port 3052. Uploads land as upload_[RANDOM].tmp in web-server\work\tomcat\localhost\local, then show up in ...\CyberPower PowerPanel Business\temp\profile\extcmd\ and run from ...\extcmd\. Script names follow pp?????.cmd, so they blend in. The report says public advisories cover hard-coded credentials, a default admin password and a file-upload path-traversal flaw in PPB. Any of those, or stolen admin credentials, could provide console access. It does not say which was used here.
- Hands-on-keyboard tradecraft: ppbed.exe → cmd.exe ran certutil to test reachability and decode base64 staged files, plus curl and a PowerShell reflective .NET load. Staging server was 84.200.205[.]145, with dt.exe, wupd.exe and hello.exe pulled from hxxp://84.200.205[.]145:55556/. wupd.exe is detected as MSIL.Killav!gen1, a .NET antivirus killer. hello.exe only prints a greeting and served as a test of the reflective-load path.
- AD collection by trial and error: the operator iterated scripts (ppskbcx.cmd, ifm.cmd through ifm4.cmd, ppotuff.cmd, ppinthw.cmd), first using ntdsutil ifm, then adding logging and done-marker files. They finally pivoted to esentutl /vss to copy the locked ntds.dit from a shadow copy, and saved the SYSTEM hive with reg save. Some revisions reused a name (ifm.cmd has three hashes), so the extcmd folder is only a partial development log.
- Implant and scope: PowerShell launched upnpcont.exe from C:\Windows\System32\config\systemprofile\AppData\Local\50679f292a1\, which side-loads an unknown atl.dll. Later vssadmin and robocopy activity kept the lineage vssadmin.exe ← powershell.exe ← upnpcont.exe ← powershell.exe ← cmd.exe ← ppbed.exe ← services.exe. Symantec did not attribute the activity and saw no exfiltration or ransomware, so intent is unclear.
DFIR takeaway: hunt on parent process, not tool name. Look for ppbed.exe as an ancestor of cmd.exe, powershell.exe, certutil, ntdsutil, esentutl, vssadmin or robocopy. Audit extcmd and temp\profile\extcmd on every PPB host against product defaults, and treat unknown scripts as hostile. Preserve those scripts, the Tomcat work directory temp files, and the upnpcont.exe/atl.dll pair. Review the console's event-to-command assignments, and check who reached port 3052.
If you find this, assume AD credentials are exposed and plan a krbtgt reset and wider credential rotation. Restrict 3052 to named admin hosts and change the default admin password. File hashes for the scripts, tools, implant and DLL are in the report's IOC section.
#DFIR_Radar
1
117
Source: security.com/threat-intellig…
38
Talos tracks UAT-12643, a suspected North Korea-aligned crypto and gaming stealer crew. Its Go RAT pulls C2 URLs from Solana memos and survives "Reset this PC" through WinRE.
Attribution is low confidence, based on TTP overlap with known North Korean 🇰🇵 APTs. Overlord RAT was seen earlier with UNK_DeadDrop, and Solana memo C2 hosting was first reported in March 2026 with Lazarus. Others also use that technique, so treat the link cautiously. Targets look like crypto holders, gamers, finance staff, and VoIP users, with keyword harvesting in English, French 🇫🇷, Spanish 🇪🇸, and Portuguese 🇵🇹 pointing at Europe.
- Staging and delivery: an opsec failure exposed the root folder of 41.216.188[.]11, a Flask file-drop on port 8000. Other ports: 5173 (Overlord WebSocket C2), 5656 and 5959 (Python stealer TCP C2). Initial access uses a spoofed PDF LNK abusing MpCmdRun.exe -DownloadFile (payload saved as RuntimeBroker.exe), a certutil -urlcache -f LNK, an HTA run by mshta.exe that writes %TEMP%\s.exe via ADODB[.]Stream, and an MSBuild project dl.xml that compiles C# in memory with a spoofed "Microsoft BITS/7.8" User-Agent. Both LNKs were built on host win-8oa3ccqae4d. The report does not say how the files reach victims.
- Overlord RAT: it requires admin, takes SeDebugPrivilege, and calls NtSetInformationProcess to set ProcessBreakOnTermination. Killing it triggers a CRITICAL_PROCESS_DIED BSoD. It has 35 commands including keylogging, webcam and mic capture, reverse SOCKS proxy, and reflective DLL plugin loading. Persistence is an HKCU autorun with a randomized value name pointing into a Network folder under local AppData, plus a WinRE hook.
- WinRE persistence is the sticky part: it writes a 20-character random-named copy to %TARGETOSDRIVE%\Recovery\OEM\, adds a .bat stub running cmd.exe /c start against that path, backs up and then creates or patches ResetConfig.xml. Any "Reset this PC" runs the implant before the OS is restored.
- Solana C2 discovery: the actor posts encrypted C2 URLs as Solana Memo Program transactions (up to 566 bytes). The RAT calls getSignatures, then getTransaction with jsonParsed encoding, and only accepts instructions whose data is exactly 43 characters, the length of a Base58 WebSocket URL after AES-256 encryption. The actor rotates C2 by posting a new memo, with no redeployment.
- Good Game stealer (public PyInstaller tool by CirqueiraDev, Python 3.12.0) and clipboard hijacker: Good Game runs a Notepad-style tkinter decoy and sandbox checks (VM MAC OUIs, under 2 GB RAM, uptime under 5 minutes, wineboot.exe). It persists with an HKCU Run value named Zoipers and a Startup folder copy, and sets DisableTaskMgr=1 under HKLM\...\Policies\System during the run. It takes browser data, Discord and Roblox tokens, 21 wallets, game launchers, and keyword-filtered files. Staging artifacts: SK_ plus 8 random characters ZIP in temp, url_opened_<hash>.flag, a file named skipped. Exfil goes to upload.gofile[.]io, then a Discord webhook alert with hostname, username, and OS. The clipper uses mutex UuuRjinBZPBMIjnc, copies to %APPDATA%\MsgEdgeUpdate.exe (hidden and system attributes), adds a Run value MsgEdgeUpdate, and polls the clipboard every 500 ms to swap detected wallet addresses.
Takeaway: do not kill a suspected Overlord process on a live host. Isolate it and capture memory first, since termination crashes the system. On disk, check Recovery\OEM and ResetConfig.xml, because rebuilding via Reset this PC will not evict it. Hunt for MpCmdRun.exe -DownloadFile, certutil -urlcache, mshta.exe or msbuild.exe pulling remote payloads, the Run values and the MsgEdgeUpdate.exe path above, and endpoint traffic to Solana RPC, Discord webhooks, or gofile. IOCs are in Talos's GitHub repository linked from the report.
#DFIR_Radar
1
307
Source: blog.talosintelligence.com/u…
34
Calif says Java4Shell (CVE-2026-114938, CVSS 9.2) enables pre-auth RCE in Jira, Jenkins, Keycloak, with no known exploitation yet. Hunt hs_err_pid crash logs and IllegalArgumentException bursts from java.util.Base64$Decoder.
#DFIR_Radar
1
153
NVISO shows how to catch base32 DNS tunneling with Sentinel Summary Rules on Microsoft DNS Analytical events, cutting DNS ingestion cost by roughly 95% while keeping alerting intact.
- The design: all Microsoft DNS Analytical events go to the Data Lake. A Summary Rule runs the detection logic against that table and writes only the hits to an Analytic Tier table (dns_sr_alerts_CL). Unlike the firewall part of the series, which aggregated records, this stores alerts rather than DNS records. That is why savings beat the previous post. Each Summary Rule run also adds a query cost, which NVISO describes as relatively low.
- The technique: DNS tunneling encodes host data (credentials, files, host or network info) into subdomains of an attacker-controlled domain, with base32 as one example encoding. The same channel carries commands back. Sliver C2 is named as a tool that uses base32 for DNS C2. NVISO's lab emulation is a PowerShell loop that calls Resolve-DnsName for TXT records on random 45-character base32 labels under a non-existent domain, so every response is NXDOMAIN.
- The KQL logic: it reads ASimDnsActivityLogs where EventProduct is DNS Server, EventSubType is request or response, and SrcIpAddr is private. It regex-matches DnsQuery for a base32 label of at least 24 characters (a-z2-7 or A-Z2-7) followed by two or more further labels. Results group by SrcIpAddr and second-level domain. It fires when Count is at least 50 and distinct subdomains are at least 10, so many unique subdomains under one domain from one host is the signal.
- Alert plumbing: the query sets alert_title, alert_description, alert_severity (Medium) and alert_notification (dns_sr_alert). One Analytic Rule filters dns_sr_alerts_CL on isnotempty(alert_title) and alert_notification == "dns_sr_alert". That lets you add more Summary Rules later without building a new Analytic Rule each time. In the lab, the output showed the test host 192.168.20[.]105 talking to the internal DNS server 192.168.20[.]101.
- Scheduling and scope: the report recommends a 1-hour run interval so threshold rules have enough records to trip. Thresholds need a per-environment baseline, and lowering them makes the rule more sensitive. The query is written for Microsoft DNS Analytical events. Other DNS products need field names and schema adapted.
Takeaway: the full KQL for the Summary Rule and Analytic Rule is in the report. Copy it, then tune the 50 request and 10 subdomain thresholds against your own baseline. Because the alert table keeps only the hits, your raw DNS evidence for scoping and timelines stays in the Data Lake. When an alert fires, pivot there on SrcIpAddr and the second-level domain to rebuild the full query history for that host.
#DFIR_Radar
1
1
173
Source: blog.nviso.eu/2026/11/17/red…
34
AuditTeam sells extortion as a paid security audit: its "Audit & Consulting fee" supposedly buys a vulnerability report instead of a data leak. The report says the group runs a custom implant framework, OpsAgent, that was rewritten from C++ to Rust during the campaign.
- Scope: the campaign started January 2026 and ran through at least October 2026. Victims were found in Russia 🇷🇺, Italy 🇮🇹, the Philippines 🇵🇭, Senegal, Finland 🇫🇮, China 🇨🇳 and Hong Kong. The leak site claims victims in seven more countries. Researchers first saw OpsAgent in Hong Kong on March 10, 2026. The ransom note never mentions encryption, but the cryptor module drops it, so if a victim sees the note, files are already encrypted.
- Loaders and injection: the C loader (April) injects shellcode into msiexec.exe with CreateProcessA, VirtualAllocEx, WriteProcessMemory and CreateRemoteThread. The .NET loader (June) uses NtQueueApcThread. The Rust loader (September) pulls an encrypted payload from C2 and runs the Execute export inside its own process, with no remote injection. Implant version 5.2.8 pushes mimikatz_spawn, execute_assembly, pth_spawn and runpe as shellcode over a named pipe, XOR-encrypted, into msiexec.exe by default. The monitor command uses notepad.exe instead. A DbgInjector mode starts a process in a debug state, sets a hardware breakpoint on FileTimeToSystemTime and redirects RIP without creating a thread.
- C2 discovery: the implant can read a public GitHub repo, hxxps://github[.]com/RonaldCMalcolm1983/staticPage, for an AES-encrypted C2 address wrapped in GetConfig and EndConfig markers. C2 sits on CloudFront subdomains such as d19v1tntmz2i98.cloudfront[.]net and d35htktbyt7nxc.cloudfront[.]net, so IP blocking is impractical. One standalone IP is 157.20.104[.]111 (AS 136557). The Rust build starts with HTTP POST, then upgrades to WebSocket over TLS. The repo was removed in November 2026, but the latest Rust implants ship with the GitHub feature disabled, so removal does not mean the campaign is over.
- EDR killer: a driver disguised as Beijing Jiangmin Antivirus (KV series) with a valid signature talks over the \KvCorePortX0 filter port. Command 0x1D plus a filename kills that process. The kill list has about 220 entries, covering Defender, Sysmon, splunkd, wazuh-agent, TeamViewer, Veeam, Acronis and major databases.
- Cryptor and anti-recovery: AES-CBC or ChaCha20 per file, with the file key protected by X25519. It kills SQL Server, Oracle, MySQL, PostgreSQL, MongoDB and Redis and stops Veeam. It also deletes shadow copies, empties the recycle bin and wipes temp folders. Deleted files are overwritten with random data first, so expect little carving success.
- Artifacts: the report lists syscache.dat under C:\ProgramData\Microsoft\Windows\Caches, user AppData Local Microsoft\Windows\Caches and Local Microsoft\CLR_v4.0, the implant's module directory and %TEMP%, plus C:\Users\Public\kv_dbg.log and kvcore.sys in System32\drivers and Sysnative\drivers. The driver hash (listed as a legitimate AV driver) is 109cdc24ccd4872a3e4bf6f2d4f3c3d4.
Takeaway: hunt for kvcore.sys loading from the drivers folder, kv_dbg.log in C:\Users\Public, and syscache.dat in those Caches, CLR, module and temp paths. Look for msiexec.exe or notepad.exe holding injected memory or named pipes with no matching install or user activity. Check for non-browser processes reaching the GitHub repo or CloudFront subdomains. Capture memory and a triage image before the cryptor runs, because shadow copies and deleted files will not help afterward. Hashes, paths and domains are in the report's IOC section.
#DFIR_Radar
1
2
269
Source: securelist.com/auditteam-ops…
34











