Patch Dell RecoverPoint Today, Then Rethink Your Whole Trust Model

The cover story

CVE-2026-22769 is a hardcoded admin credential in Dell RecoverPoint for Virtual Machines, sitting in tomcat-users.xml on the appliance. CVSS 10.0. Unauthenticated WAR deploy to /manager/text/deploy gets you root. Mandiant and Google's GTIG documented active exploitation by UNC6201, dropping a new backdoor called GRIMBOLT — compiled as native AOT C# specifically to strip .NET metadata and neuter every standard reverse-engineering tool. Persistence lives in a hijacked convert_hosts.sh fired from rc.local. Lateral movement pivots through VMware via phantom vNICs and iptables Single Packet Authorization on vCenter appliances.

What happened. A nation-state actor is inside backup infrastructure at customers who haven't patched yet. Auditors should be grepping /home/kos/auditlog/fapi_cl_audit_log.log for /manager hits right now.

Why it matters. RecoverPoint is the thing you fall back to when everything else is on fire. UNC6201 isn't there to exfiltrate — they're there to deny you recovery. Combine that with the second story of the day and it stops being a patch and starts being a pattern.

What to do. Patch RecoverPoint VMs today, deploy Mandiant's GRIMBOLT YARA across VMware estates, hash-check convert_hosts.sh against a known-good baseline, and sweep vCenter for Ghost NICs and unexpected iptables rules. Deadline: 72 hours. Sources: Mandiant advisory, GTIG bulletin.

Yes, but — a defender could reasonably say this is one vendor, one CVE, and patch cycles handle it. The answer is that a hardcoded credential in Tomcat shipped in an EMC-lineage DR product is a story about how the industry ships critical infrastructure, not an outlier. And UNC6201's tradecraft evolution — BRICKSTORM to GRIMBOLT, native AOT, VMware pivots — is the part that generalizes.

The EDR blind spot nobody's talking about

ADWSDomainDump enumerates Active Directory over ADWS on port 9389 instead of LDAP. It bypasses both CrowdStrike Falcon and Microsoft Defender for Endpoint. This is not a signature gap. The vendors don't have ADWS as a data source. Until they do, your reconnaissance-phase telemetry is a fiction.

The tool is public. Assume it's already in the operator toolbelt of the same actors running the GRIMBOLT campaign. Restrict 9389 to admin subnets this week, then instrument network-layer detection on the port — volume, source workstation, query cadence. If your SIEM can't do it, your firewall can.

The password manager story is worse than it reads

ETH Zurich's Applied Cryptography Group demonstrated 25 attacks across Bitwarden (12), LastPass (7), and Dashlane (6). Roughly 60M users. The attacks work by impersonating the sync server during routine operations — not a full breach, just a compromised sync path. Root cause is a mix of feature bloat and 1990s-era primitives. The full paper drops at USENIX Security 2026, which is the exact window between "credible academic finding" and "weaponized tooling on GitHub."

Don't rip out your password manager this week. Do commission the assessment now so you're not doing it under a headline in six months. The interesting move is treating vault sync as an untrusted channel and layering hardware keys where the blast radius warrants it.

The pattern under the day

The three security stories converge, and the AI stories converge somewhere else, and both convergences point at the same operator lesson: you are paying a complexity tax you never budgeted for.

On the AI side, three independent findings landed the same punch. FloTorch's 2026 RAG benchmark showed naive 512-token recursive chunking beating semantic and proposition-based approaches on accuracy at 3-5x fewer vectors. LangChain jumped Top 30 to Top 5 on Terminal Bench 2.0 by changing the harness — self-verification, structured tracing — with no model swap. And a first-principles walk through transformer inference math shows the H100 running at roughly 1% utilization during decode because arithmetic intensity collapses to 1 FLOP/byte against a 295 FLOP/byte threshold. The memory wall isn't an engineering problem. It's physics that gets worse each GPU generation.

Context length is where this bites your P&L. A 7B model serves 278 concurrent users per H100 at 4K context and 8 at 128K. That's a 35x per-user cost swing hidden inside what your PM thinks is a feature flag. Every agent that passes raw trace history between hops walks you up that curve.

So: simple RAG beats fancy RAG, harness engineering beats model swaps, and the biggest lever on unit economics is the context budget you set in a config file. The teams shipping the best AI features this quarter aren't the ones with the newest model — they're the ones who profiled what they already have.

What to actually do this week

Four moves, in order of blast radius.

One: patch RecoverPoint and hunt for GRIMBOLT indicators. This is the drop-everything item. If you don't run RecoverPoint, treat it as a rehearsal — the tradecraft transfers.

Two: put ADWS on port 9389 behind network segmentation and add a detection rule keyed on volume and source. Your EDR won't do this for you.

Three: instrument context length in production and correlate it with COGS. If your median context is above 8K and you don't know why, you have a hidden $0.30/hour-per-user problem masquerading as a product decision.

Four: run a two-week A/B on 512-token recursive chunking against whatever semantic pipeline you built last quarter. If simple wins — and the benchmark says it will — you just found a 3-5x infrastructure cost cut with zero accuracy loss. That's a bigger quarter than most model upgrades will get you.