Engineering & Technical
The Engineer
PoeLLM finds its command server by reading four words from a poem on GitHub.
A hard-coded dictionary maps those words to the C2 address at runtime. That leaves no domain for you to blocklist and no date-seeded pattern for DGA detection to learn. The only fetch on the wire goes to the site hosting the poem, and the backdoor has held 3,400+ self-hosted open-source AI servers since April.
In Play
FortiBleed and PoeLLM Hit Edge and AI Infrastructure
Two live campaigns hit infrastructure you run. CyberScoop reports FortiBleed, an FBI/Secret-Service-flagged credential-abuse campaign, is targeting 400,000+ Fortinet firewalls and VPN gateways. PoeLLM has backdoored 3,400+ self-hosted open-source AI servers since April. If you run edge gear or Ollama/vLLM nodes, both are live right now; mechanics and actions are in the security deep dive.
Ask ClarityOpen-Weight Models Reach Mythos-Class Cyberattacks
Frontier-class offensive cyber capability is moving into open weights that no vendor can switch off. That puts attacker and defender at the same capability. This is a forward-looking risk, separate from today's live campaigns. The evidence is covered in the security deep dive.
Ask ClarityGeneration Got Cheap; Verification Is the Bottleneck
Generation now outruns review in frontier math, in chip design and in your own agent pull requests. Where generation scales, verification coverage becomes the metric that matters. The verification deep dive has the evidence and actions.
Ask ClarityNetflix Uses an LLM as a Ranker That Never Generates Text
Netflix's GenRec uses an LLM as an encoder that scores a closed catalog and never generates text. The pattern, not the lift, is the transferable lesson for any LLM ranking or labeling task. The architecture and numbers are in the GenRec deep dive.
Ask ClarityTexas Froze Switch-On for Big Data Centers
a16z reports ERCOT has paused energizing every data center of 75 MW or more, including 17 that cleared every other step, with the Batch Zero fix stalled until a December audit and air permits frozen too. The large-load queue grew from 63 GW at end-2024 to 474 GW by June 2026. New capacity arrives on curtailable terms: a 30-minute full-load shed, or a 1-minute cutback if your own plant trips. Any 2027–28 GPU capacity sited in ERCOT now carries schedule and curtailment risk.
Ask Clarity
Deep Dives
- ●
FortiBleed, PoeLLM, and the Open-Weight Risk Coming Next
Two live campaigns map onto infrastructure you run. Separately, offensive capability is starting to ship as weights no vendor can switch off. Together they mean detection posture, not patch count, decides your exposure.
The C2 mechanic that breaks blocklists The best engineering in today's intel is PoeLLM's command-and-control . It uses no hard-coded domain and no date-seeded generation algorithm. The implant reads a poem posted on GitHub , pulls four specific words, maps…
4 action items
- ●
Verification Is the Bottleneck, and the Verifier Is the Product
Five independent stories describe one dynamic — generation now outruns review — and the teams that win treat the deterministic checker, not the model, as the thing worth building.
Start with what works in production The clearest template is GPT-Synopsys , OpenAI's chip-design agent. TLDR Hardware strips the branding. Underneath is a verifier-gated loop (objective, run tool, interpret output, modify, re-run) that stops only when verification passes. It is…
3 action items
- ●
Netflix's GenRec: Use the LLM as an Encoder, Not a Generator
The part worth copying isn't the microscopic engagement lift — it's a serving architecture that makes hallucination structurally impossible and ties cost to input tokens you control.
The one design choice that matters GenRec is a ranker. It reads a plain-text description of a member's behavior, scores the catalog, and stops after prefill . A generative LLM runs prefill and then decodes output tokens one at a…
2 action items
The edition continues
Take the signal into the room.
Sign up or log in to read all 3 deep dives in full, plus the final take.
Read the full editionContinue with LinkedIn