The AI Briefing

Version 2.0 of the open-source terminal coding agent, 493 points and 200 comments within hours of the announcement.

A 24B-parameter sparse model with roughly 2B active parameters, built for agent workloads running locally — tool calling without the cloud round-trip.

Experts tell CSO Online the 'rogue agent' incidents reflect containment and monitoring failures at frontier labs, not autonomous intent. 869 points, 504 comments.

Ground infrastructure for satellite broadband is now an attack surface; 344 points and 223 comments.

Android users report automatic phone-number verification propagating across every account on the device, with no obvious opt-out. 445 points.

The financing behind the compute buildout is being sold on quickly — a read on how durable lenders think the demand is. 357 points.

Open-source vector index for Postgres positioned as a drop-in alternative to the two standard pgvector index types.

An essay against the 'just prompt it' framing; 768 points and 593 comments of argument about where the skill actually lives.

A point release carrying experimental features aimed at dreamers; 258 points, 85 comments.

A design analysis traces the lineage of Anthropic's Scion back to Mononoke; 342 points and 128 comments.

Agent benchmarks measure task success under a fixed harness, so nobody tests the thing that actually determines results — the scaffolding around the model. Doc2Harness makes harness construction the task: a build model receives an interface skeleton plus a development document distilled from twelve frontier harnesses and must produce a working harness, which is then frozen and driven end-to-end by scripted mock models in isolated containers. It is the first benchmark to score the engineering layer directly rather than assume it. [coding-agents, harnesses, benchmarking]

Passing the tests an agent was given proves little — the open question is whether agents can build their own checks that faithfully capture a specification. SpecGuardBench poses 291 tasks from 13 real repositories across 8 verification-logic types, requiring agents to construct executable validation harnesses that reject spec-violating implementations. It measures whether generated evidence is trustworthy, not merely whether code runs. [coding-agents, verification, benchmarking]

Long coding sessions force context compaction, and the usual model-written summary silently drops facts and commitments the agent still needs — worst for the small models now running locally. Transparent-compact, built for the open-source Pi agent, is deterministic: user and assistant prose is guaranteed unchanged, and stripped information is explicitly flagged rather than paraphrased away. It targets exactly the lightweight, high-volume deployments where a lossy summary hurts most. [context-management, coding-agents, lightweight-models]

Agentic systems assume a single developer controlling every agent, which caps scale at one person's budget. This paper models multi-developer agentic systems where specialists' agents compose: on GAIA, a population of specialists each capped at 35% accuracy reaches 70% when combined — 12 points above a median general-purpose system built with several times each specialist's budget. Composition, not capability per agent, does the work. [multi-agent, orchestration, scaling]

Model-level privacy benchmarks miss the real risk: agents with tool access moving sensitive data across email, documents and calendars in ways that violate context-specific disclosure norms. AgentCIScope turns public and synthetic privacy cases into executable multi-step agent tasks in realistic application environments and scores whether the resulting information flows preserve contextual integrity. It evaluates the flows, not the model's refusal behaviour. [privacy, agents, benchmarking]

Existing rule-reasoning benchmarks hand the model a handful of tidy facts, which flatters real performance where domain rules are scattered across many files. RuleTrek builds multi-domain scenarios from structured rule graphs, composing sub-scenarios incrementally while guaranteeing consistency and solvability, and tests rule identification, application and conflict handling in interactive workspaces. Closer to the compliance-shaped mess agents actually face. [agents, rule-reasoning, benchmarking]

Frames attacks against three nested attacker levels along the evidence path — seed chunks and the graph evidence assembled from them — an underexamined surface now given a formal framework.

Constructs MMKGs with explicit guidance rather than the usual unguided pipelines, targeting multimodal GraphRAG for MLLMs — quality of the constructed graph is the bottleneck it attacks.

Technical language processing over the largest written record of how industrial equipment fails — abbreviated, misspelt, mostly-empty-field work orders — turned into agent-accessible failure knowledge.

Show HN: compiles Ossie models straight to SQL, no serving layer in between. Early — 5 points, 2 comments — but the zero-server angle is the interesting bit.

The Go release blog details changes to how net/http reuses HTTP/1 connections — relevant to anyone running high-fanout clients.

Cloudflare's data shows bot traffic migrating to other infrastructure during the outage instead of disappearing — the botnet outlived the site. 248 points.

Oxide publishes a way to run a complete rack — the sleds, switches and all — in emulation; 107 points and 10 comments.

An Android NDK issue documents the gap between shipped binaries and published source; 491 points and 242 comments.

A first-person guide through diagnosis and treatment; the top story on the front page at 1,131 points.

An investigation into Amazon's pricing influence beyond its own storefront; 787 points, 620 comments.

Apple's new home product page; 752 points and 903 comments, mostly argument.

Sans-serif and slab-serif monospace families, endlessly customisable; 714 points.

A critique of the genre's length norms, posted to an audience largely composed of the genre's authors; 300 points, 222 comments.