RepoJournal
Local LLMs Local LLMs
69 wires and counting

$ follow Local LLMs

Keep up with Local LLMs in about 3 minutes: what actually shipped — the commits, pull requests, releases, and security advisories that matter.

or

fair warning: these emails are deeply technical. diffs, version numbers, CVEs, benchmark deltas. if that's not your idea of a good read, this isn't your newsletter.

Folds into your digest — weekly by default, monthly if you prefer. Unsubscribe in one click.

$ status

wire 2026-08-13
stories 249
RSS Archive Reviews Topics

© 2026 RepoJournal Home Showcase Explore How it works Privacy

$ the-wire · showcase

OLLAMA PATCHES CRITICAL SSRF VULNERABILITY IN MANIFEST HANDLING

By RepoJournal · Filed · About Local LLMs · Composed from the cited sources · methodology

ParthSareen Vignesh parabelboi Chipmunk Daniel Bevenius yifjiang Akashcodes732
17 people shipped this
ParthSareen @ParthSareen 2 cited Vignesh @vigneshakaviki 1 cited parabelboi @parabelboi 1 cited Chipmunk @CHIPMUNK-T0T 1 cited Daniel Bevenius @danbev 1 cited yifjiang @yifjiang 1 cited Akashcodes732 @Akashcodes732 1 cited bigPYJ1151 @bigPYJ1151 1 cited yctseng0211 @yctseng0211 1 cited mmangkad @mmangkad 1 cited michaelzhang-ai @michaelzhang-ai 1 cited jthomson04 @jthomson04 1 cited Eve @netrunnereve Aleksander Grygier @allozaur ovidiusm @ovidiusm yewentao256 @yewentao256 ajith-sirra-amd @ajith-sirra-amd

A rogue OCI registry could exploit duplicate digest collisions in Ollama's blob verification to exfiltrate data from internal endpoints.

Ollama shipped a critical fix [1] addressing a skipVerify map collision that skips hash verification when a manifest's config and layer share the same digest. The vulnerability allows an attacker to serve a malicious manifest and redirect blob downloads to internal endpoints, bypassing Ollama's verification entirely. This is a production-grade SSRF risk that demands immediate patching.

On the feature front, Ollama's agent tools got a major upgrade [2] with support for multiple edits per call, letting models apply batch file replacements against the original content with indexed error reporting for self-correction. The OpenAI Responses API now supports web search [3], enabling Codex and other models to execute searches through Ollama and receive results inline. Llama.cpp added a read_media tool [4] that lets vision models analyze server-side images, and shipped multimodal slot state save/restore [5] so you can persist and restore sessions with image inputs intact. Semantic versioning is now live [6], bringing stability to release tracking and CI workflows.

vLLM fixed KV cache reporting to correctly identify FULL_ATTENTION specs [7] instead of marking them UNKNOWN, added PowerPC/VSX unquantized MoE support [8], and rolled out registry-based layer caching for x86 CPU image builds [9] to eliminate redundant recompilation. SGLang swapped AMD's PR gate to ROCm 7.2 [10], demoting 7.0 to shadow testing, reverted a Kimi K3 fusion that broke CI [11], fixed multimodal partition failures on 2-GPU AMD shards [12], and shipped cache salt support for KV events [13] to prevent external routers from colliding with caller-defined namespaces. Activity: 123 commits, 122 PRs, 4 releases across the stack.

Action items

References

  1. [1] server/images: prevent skipVerify map collision with duplicate digests (#15504) ↗ ollama/ollama
  2. [2] agent: allow multiple edits per edit tool call ↗ ollama/ollama
  3. [3] openai: support web search in Responses API ↗ ollama/ollama
  4. [4] ui: add read_media tool (#25877) ggml-org/llama.cpp ↗
  5. [5] server : support slot save/restore with media inputs (#26640) ggml-org/llama.cpp ↗
  6. [6] cmake : introduce semantic versioning (#26839) ↗ ggml-org/llama.cpp
  7. [7] [Bugfix] Report FULL_ATTENTION for uniform-base UniformTypeKVCacheSpecs groups instead of UNKNOWN ↗ vllm-project/vllm
  8. [8] [Hardware][Power] Unqualized MoE Backend for Power (VSX) ↗ vllm-project/vllm
  9. [9] [CI] Add registry layer cache to x86 CPU image build ↗ vllm-project/vllm
  10. [10] [AMD][CI] Swap the AMD PR gate to ROCm 7.2 and demote ROCm 7.0 to a daily shadow ↗ sgl-project/sglang
  11. [11] Revert "[Kimi K3] Fuse MLA gate projection into QKV-A GEMM" ↗ sgl-project/sglang
  12. [12] [AMD][CI] CI: fix AMD 2-GPU multimodal-gen partition-count abort ↗ sgl-project/sglang
  13. [13] feat: add cache salt support to KV cache events ↗ sgl-project/sglang

$ ls local-llm/reviews/ # the step-back reads

the week in review Five engine-fatal flaws hit vLLM, llama.cpp ships v0.5.0 the month in review vLLM 0.30.0 lands and nineteen engine-fatal flaws close
Topics: Local LLMs AI / ML Full archive → Weekly reviews →

Quick answers

What shipped in Local LLMs on August 13, 2026?
A rogue OCI registry could exploit duplicate digest collisions in Ollama's blob verification to exfiltrate data from internal endpoints. In total, 123 commits, 122 pull requests, and 4 releases landed.
Who contributed to Local LLMs on August 13, 2026?
17 developers shipped this update, including Vignesh, ParthSareen, parabelboi, Chipmunk, Daniel Bevenius, Eve, Aleksander Grygier, and ovidiusm, and 9 more.
What were the notable Local LLMs updates?
server/images: prevent skipVerify map collision with duplicate digests (#15504), agent: allow multiple edits per edit tool call, and openai: support web search in Responses API.
♥

$ log Local LLMs --recent

Ollama gates System One on declared capabilities, llama.cpp fixes a stale-UI trap 298 shipped vLLM's rejection path desyncs multimodal caches 285 shipped llama.cpp unblocks chunked reranking, vLLM patches three engine-core holes 221 shipped vLLM locks down multimodal kwargs, llama.cpp cleans up failed restores 136 shipped

$ log --all # elsewhere on the wire

@openclaw openclaw 2026.9.7 hardens update rollback, and Peekaboo ships GPT-6.1 Sol Agentic Coding Enterprise permission defaults get teeth as Codex, Gemini CLI and Goose ship fixes

$ ls showcases/

Rails Django Laravel Node.js Go Vercel Vue.js Kubernetes HashiCorp Arch Linux Linux Anthropic

$ init your-desk

RepoJournal can write one of these for any repo you care about. Sign in with GitHub, pick your repos and sources, and your desk gets a wire like this one every morning.