Engineering & Technical
The Engineer
The TeamPCP arrests ended the worm, not the 500,000 credentials it already took.
The propagation path ran through Trivy, KICS, LiteLLM and Telnyx. Those four sit inside CI, holding registry tokens, cloud keys and model-provider PATs. Roughly 78,000 of those secrets from 2,200 organizations were published in July, which makes the remaining question how wide your rotation has to go.
In Play
Build Runners Were The Target, Not The Packages
The Australian Federal Police arrested two alleged TeamPCP operators in Perth, but the crew's self-propagating worm had already compromised Trivy, KICS, LiteLLM and Telnyx and harvested more than 500,000 credentials, per Risky.Biz. Those are CI-resident scanners plus a proxy that concentrates every model-provider key in one process, so the loot was registry tokens, cloud keys and PATs. About 78,000 secrets from roughly 2,200 organizations already leaked publicly in July. That makes this a retrospective rotation problem, not a patch.
Ask ClarityExploited Bugs That Carry No Severity Score
CISA added nine actively exploited CVEs between August 18 and 25, including Gitea CVE-2026-60004 at CVSS 9.8 and an MLflow tracking-server SSRF, per SANS reporting. Gitea turns plain repository write access into code execution as the service account by planting an executable Git hook, and upgrading does not remove hooks already written to disk. Several of those entries carry no NVD score at all, so a severity-sorted remediation queue ranks confirmed exploitation below a medium dependency warning.
Ask ClarityPrefix Caches Miss On Chunk Order
Anthropic's prompt cache walks backward a maximum of 20 blocks looking for an earlier write, so pushing more conversation than that between two calls stops hits entirely — no error, no warning header, just a bigger bill. vLLM keys prefix blocks by a hash chain over token IDs, which makes matching order-sensitive by construction: two RAG requests retrieving the identical document set in a different order share exactly zero blocks. Your single hit-rate dashboard mixes prefill savings with a semantic cache that can return a wrong answer with a 200.
Ask ClarityWrite Cost Decided GraphRAG's Fate
LinkedIn replaced GraphRAG in its production hiring assistant with a hierarchical summarization tree because incremental update cost, not retrieval quality, was the binding constraint. Absorbing one new user signal in a graph means entity resolution plus edge re-linking across an unbounded neighborhood, while a tree re-summarizes a single path root-ward. Benchmarks published in the same window put RocksDB write amplification at 4.56x for random inserts versus effectively none for sequential keys — the same lesson at the storage layer.
Ask ClaritySNI-Based Egress Policy Loses Its Signal
Google is shipping Encrypted ClientHello at the OS level in Android 17, covering all device traffic, per Risky.Biz. Any egress allowlist, DLP rule or CI network guardrail that reads the plaintext SNI field to decide what to permit loses that input, and it will not raise an error — it will simply stop seeing destinations. The named migration paths are explicit forward-proxy CONNECT logging or DNS-layer policy with pinned resolvers, and doing it on your schedule beats discovering the gap in an incident review.
Ask Clarity
Deep Dives
- ●
Credential Density Was The Targeting Function
Signature checks, SBOM diffing and vulnerability scanners all passed clean through this campaign, which is why rotation scope — not patch speed — is the decision in front of you.
Risky.Biz places the compromised releases of Trivy, KICS, LiteLLM and Telnyx in a January-to-August 2026 exposure window. The Hacker News dates the Trivy and Checkmarx compromise to March 2026 . That disagreement is the rotation window, and the two accounts…
3 action items
- ●
Four Caches, One Dashboard, One That Can Be Wrong
Three of the caches in your inference path can only cost money; the fourth silently changes answers — and your RAG layer is probably defeating the cheap one for free.
The mechanism that makes prefix caching fragile decides where the prompt-assembly boundary goes. vLLM stores KV in fixed 16-token blocks , each identified by a hash over the parent block's hash plus its own token IDs. That chaining turns a…
3 action items
- ●
LinkedIn Cut GraphRAG On The Write Path, Not The Read Path
Two independent accounts of agent memory land on the same conclusion: retrieval quality is not what production bills you for, and key ordering is a 4.5x lever hiding in your state store.
The architecture is the part of LinkedIn's decision worth copying. Its hiring assistant runs four memory types: conversational, episodic, semantic, procedural. Behind them sit three dedicated services: ingestion, consolidation and retrieval , each with access control, traceability and a hard…
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