Linkzine

AI-generated patches slow Linux ARM64; AWS bets big on agentic everything

Thorsten

AI-generated patches are backing up Will Deacon’s ARM64 review queue, AWS is shipping agents that act before asking permission, and the LeadDev piece quietly explains why your deployment metrics no longer mean what you think they mean.

// SECURITY FOCUS

AI agents as confused deputies: how prompt injection took 20k Instagram accounts

The Meta case is a concrete example of what ‘confused deputy’ looks like in production: the agent had ambient authority it never needed, and no authorization check between ‘user asked’ and ‘agent acted’. If your agents inherit service-account credentials or session tokens from the calling user, your fleet has the same shape of exposure right now – most agent scaffolding doesn’t enforce least-privilege by default.

What to do: Audit what credentials your deployed agents inherit at runtime and add an explicit authorization layer between agent intent and privileged action.

  1. AI/LLM patches slowed ARM64 Linux 7.2 feature work, maintainer says — Phoronix · Jun 18
    ARM64 maintainer Will Deacon notes that AI/LLM-generated patches have slowed feature work for Linux 6.12’s ARM64 tree, with extra review rounds from tools like Sashiko pushing several items to the next cycle. The complaint isn’t unique to ARM64 – other upstream maintainers are absorbing the same overhead: more patches to triage, not all of them relevant or pressing. What did land for 6.12 includes new HWCAPS for the 2025 dpISA extensions, additional errata handling, groundwork to unmap kernel data and BSS sections from the linear map, and support for the MPAM v0.1 architecture. The broader signal here is that AI-assisted contribution is already measurable as a maintenance cost, not just a submission volume curiosity.
  2. Top announcements of the AWS Summit in New York, 2026 — AWS News Blog · Jun 17
    AWS Summit New York 2026 dropped a wide batch of agentic AI announcements centered on Amazon Bedrock AgentCore and a new umbrella called AWS Continuum. On the build side: AgentCore Harness is now GA for running production agents via config rather than hand-rolled orchestration loops; a managed Knowledge Base adds native data connectors and multi-step agentic retrieval; and a Web Search tool grounds agents in current web results without data egress. A coming-soon service called AWS Context promises to map organizational data into a runtime knowledge graph for agents. On the security side, AWS Continuum (gated preview) ingests findings, prioritizes by business impact, and drives fixes through existing pipelines – the Security Agent inside it adds STRIDE threat modeling and IDE integrations via Kiro and Claude Code. Other notable items: Kiro gets a native iOS app (gated preview) for monitoring and steering sessions from mobile; AWS WAF adds a Bot Control feature that lets publishers charge AI bots for API access; and S3 now supports up to 1 GB of mutable, queryable annotations per object, aimed at agent workflows. Most of the bigger items are preview or gated, so expect a gap between the announcement and production availability.
  3. The 8 software engineering metrics AI broke — LeadDev · Jun 16
    AI coding tools have broken most standard engineering metrics by severing the link between effort and output – deployment frequency, cycle time, PR volume, lines of code, and DORA change failure rate all inflate automatically as a side effect of normal agent use, without any corresponding improvement in quality. Goodhart’s Law used to require deliberate gaming; now it’s just what happens when you run a model. The article names three metrics that still hold up: time to recover (MTTR), escaped defect rate, and actual business/customer outcome metrics – because AI can’t inflate them without real consequences. The underlying argument is that most dashboards were measuring effort as a proxy for value, and now that effort is nearly free, those proxies are just measuring token spend. If your team’s metrics are trending up while your incident rate and defect escape rate stay flat or rise, the dashboard is lying to you.
  4. Navigating the future: Schiphol Airport’s journey to shift-left platform engineering — Red Hat Blog · Jun 16
    Schiphol Airport – handling 70 million passengers annually and processing roughly 1 petabyte of camera data daily – built a hybrid OpenShift platform to satisfy Dutch defense regulations requiring biometric data to stay on-premise, while still using Azure Red Hat OpenShift for other workloads. A team of seven engineers now manages 1,800 deployments and 400 namespaces, and onboarding new teams with full security enforcement dropped from weeks to hours after adopting Red Hat Advanced Cluster Security as a managed service. The airport also reported a 120% increase in cyberattacks last year, which pushed the shift-left security approach up the priority list. This is a Red Hat-authored case study, so treat the “hours not weeks” and single-week setup claims as vendor-reported rather than independently verified – but the architectural constraints around data sovereignty and the team-size-to-footprint ratio are the interesting operational details here.
  5. Introducing AWS Continuum: Security at machine speed — AWS Security · Jun 17
    AWS announced Continuum for code vulnerabilities, an agentic security product now in gated preview that aims to handle the full vulnerability lifecycle – discovery, prioritization, validation, and remediation – using multiple frontier AI models. The pitch is that traditional telemetry-and-dashboard workflows can’t keep up with AI-assisted attack tooling, so remediation itself needs to be automated. It ingests existing vulnerability backlogs alongside its own scans, confirms exploitability in a sandbox, and proposes patches that it then validates using the same exploit system. It ships in a human-in-the-loop “learn mode” first, with an “enforce mode” available once you’ve built confidence in its judgment. This is a vendor announcement with no independent benchmarks, no pricing disclosed, and the preview is gated – so the claims about false-positive reduction and patch quality are unverified so far.

// In other news

ai

  • AI Use by the US Government (Schneier on Security) · Jun 17 — Schneier flags that the Trump administration’s April OMB disclosure quietly confirmed AI is already automating federal processes at scale, with limited public visibility into which decisions are affected.
  • [AINews] GLM-5.2: the top Frontend Coding model in the world, IndexShare for Speculative Decoding (Latent Space) · Jun 17 — GLM-5.2 from Zhipu AI claims the top spot on frontend coding benchmarks among open models, and the issue also covers IndexShare, a speculative decoding technique that could cut latency for serving large open-weights models.
  • A near-autonomous AI chemist improves a challenging reaction in medicinal chemistry (OpenAI Blog) · Jun 17 — OpenAI and Molecule.one ran a near-autonomous AI chemist on a real medicinal chemistry optimization problem, with the system iterating on synthesis conditions without constant human direction — the methodology is more interesting than the headline.
  • Introducing LifeSciBench (OpenAI Blog) · Jun 17 — OpenAI releases LifeSciBench, an expert-authored benchmark for life science reasoning tasks, giving teams evaluating models for biomedical use a more domain-grounded alternative to generic science evals.
  • Local Qwen isn’t a worse Opus, it’s a different tool (Alex Ellis) · Jun 17 — Alex Ellis tests local Qwen against Claude Opus on real workloads and finds local models aren’t a drop-in replacement — useful framing for anyone deciding whether to route tasks locally vs. to a frontier API.

cloud

culture

  • CI/CD with Robert Erez (Pragmatic Engineer) · Jun 17 — Octopus Deploy’s Robert Erez covers where GitOps and progressive delivery patterns are actually landing in practice, including how AI tooling is changing CI/CD pipeline design — more operational than the usual podcast fare.
  • Fragments: June 16 (Martin Fowler) · Jun 16 — Martin Fowler’s June 16 fragments include Pragmatic Dave Thomas on sustaining a programming career long-term — short, but worth two minutes if you’re mid-career and thinking about what holds up.

iac

linux

obs

sec

  • Introducing the Red Agent POV Series (Wiz Blog) · Jun 17 — Wiz’s Red Agent series documents how their AI-driven attacker surfaces complex, chained exploit paths in real cloud environments — early episodes cover the methodology, not just the findings.

web

  • How PHP Attributes Changed the Way I Write Livewire (Freek Van der Herten) · Jun 17 — Bert De Swaef’s approach uses PHP attributes to attach validation rules, URL binding, and event listeners directly to Livewire component properties, cutting the boilerplate that piles up in larger components.
  • The artisan dev Command in Laravel 13.16.0 (Laravel News) · Jun 17 — Laravel 13.16.0 ships a new `artisan dev` command aimed at streamlining local development workflows — worth checking the changelog if you run custom `serve` scripts.
  • How We Cached Laravel News at the Edge with Fast Laravel (Laravel News) · Jun 17 — Jason McCreary walks through how Laravel News uses Cloudflare edge caching on a high-traffic Laravel app with dynamic content — concrete config decisions included, not just theory.
  • Recap: Restoring removed version history. (Make WordPress Core) · Jun 17 — WordPress core team recaps the work to restore block-editor version history that was lost during the manual Gutenberg sync process — affects anyone tracking Git blame across the 5.0-era codebase.

AI-generated patches slowing Linux, AI agents breaking metrics, AI agents breaching accounts — worth sitting with that for a day.

, , , ,
, , , ,

Leave a comment