A supply chain ransom hit Grafana’s CI runners, a Cisco SD-WAN zero-day is being used for lateral movement in production right now, and both Fedora and Red Hat published pieces about what happens when humans stop owning the security decisions in their own pipelines.

1. The AI code paradox: Moving fast without breaking security — Red Hat Blog · Jun 24
   Red Hat argues that both banning AI coding assistants and allowing unrestricted use create security problems – the former pushes developers toward shadow IT, the latter lets AI-generated code bypass review with unvalidated dependencies and privilege-escalation risks. The post frames the response as three pillars: governance policy (developers own AI-generated code, full stop), security-focused prompting skills, and AI-aware CI/CD pipelines that add automated slop filtering, SAST/DAST, and dependency scanning before code reaches a human reviewer. Red Hat has open-sourced a library of 130+ security skills for AI coding assistants and an AI code review tool to help automate the quality gate. The vulnerability load angle is worth taking seriously: FIRST projects up to 100,000 CVEs registered in 2026, more than double 2025’s count, meaning AI-generated code adds review volume on top of an already-growing advisory backlog. The concrete tooling references – prodsec-skills, LoLa, Anti Slop, RapiDAST – make this more than a framework post, though the heavy Red Hat branding means you’re reading a vendor’s preferred stack, not a neutral survey.
2. Don’t wrap OpenTelemetry — your abstraction layer is breaking the API contract — OpenTelemetry · Jun 24
   Wrapping the OpenTelemetry API in a convenience layer – an IMetric interface, a TelemetryHelper class, a MetricsWrapper module – silently destroys the performance model OTel was built around. The .NET SDK’s Histogram<T>.Record() has allocation-free overloads for 1-3 attributes and a stack-allocated TagList for up to 8; Rust’s API takes a borrowed slice. The moment your wrapper signature requires any collection type, every caller heap-allocates on every measurement. The pattern also tends to hide instrument creation inside the record call itself, which turns what should be a direct reference into a ConcurrentDictionary lookup or a Mutex-guarded HashMap acquisition on every hot-path write. Beyond the performance cost, developers learn your abstraction instead of OTel, you own an API layered on top of an API, and bugs get harder to trace. Legitimate exceptions exist – dual-write migrations, codegen-based governance tooling – but for most teams the wrapper costs more than it saves. Skip it and point people at the official docs.
3. Log management for AI workloads: How to bring your logs and telemetry plan into the AI-first century — Dynatrace Blog · Jun 24
   A Dynatrace-commissioned survey of 450 senior IT leaders finds AI workloads are breaking traditional log management assumptions: teams now average seven log tools, 50% of organizations discard an average of 86% of logs purely to control costs, 85% struggle to ingest logs at AI scale, and 80% say converting telemetry to insight is actively delaying AI projects reaching production. The core argument is that logs alone can’t explain AI behavior – you need logs, traces, metrics, and lifecycle context unified in a single queryable layer, enriched at ingest rather than reconstructed later. The action plan covers five areas: unify telemetry into one continuously queryable platform, establish causation by correlating logs with traces automatically, cut costs by dropping rigid schemas and rehydration dependencies, standardize instrumentation at ingest, and shift from reactive troubleshooting to real-time drift detection with automated remediation. The framing is vendor-neutral in principle but the recommended approach maps cleanly to Dynatrace’s own product, so read the stat sourcing with that in mind.
4. Grafana Labs post-incident review: TanStack npm supply chain ransom incident — Grafana Labs
   The TanStack “Mini Shai-Hulud” supply chain attack hit Grafana Labs on May 11 via malicious code executed on self-hosted CI runners, leaking credentials. Grafana rotated what it believed were all affected credentials, but missed one – which a threat actor used on May 14 to clone the entire repository collection before demanding a ransom on May 16. Grafana refused to pay, following FBI guidance, and confirmed no customer production systems or the Grafana Cloud platform were affected, though private repos containing internal tooling, business contact names, and some marketing email addresses were exfiltrated. The remediation was substantial: 1,500 security-focused PR reviews, audits of 280 GitHub applications, scanning of 1,200 repositories, and 2,300 PR reviews of a single critical repo. An independent Mandiant investigation completed June 18 found no evidence of code tampering or repository poisoning. Going forward, Grafana has deployed a token broker for short-lived credentials, fine-grained access controls, and is compartmentalizing GitHub organizations – steps worth comparing against your own CI/CD credential rotation and scoping practices.
5. Fedora moves to require 2FA for packagers after alleged account compromise — LWN.net · Jun 24
   After a compromised account let an AI agent cause downstream problems for the project, Fedora is starting mandatory 2FA with packagers first — LWN covers the discussion, the rollout plan, and the friction points that make MFA harder in contributor-heavy open source projects than it sounds.

---

// In other news

Rotate your CI credentials, enable 2FA on your package accounts, and patch the SD-WAN gear — then read the OTel wrapping post. Back tomorrow.