AI Agents

AgentXploit splits pre-deployment agent auditing into two roles: an Analyzer that traces attacker-controlled inputs to sensitive operations across a repo, and an Exploiter that turns those paths into runtime attacks confirmed by an external verifier. On its own AgentXploit-Bench (72 reproducible vulnerabilities across 12 agent systems) it hits 59.3% end-to-end success versus 38.4% for Codex, and 46.3% under matched token budgets. On AgentDojo the Exploiter alone reaches 79.2%. The reproducible benchmark is the standout contribution.

MetaPermit decouples semantic inference from security enforcement for agent tool access. Instead of enumerating user intents in static policy, it derives a compact set of task-independent meta-attributes capturing intent, execution context, and the proposed call; at runtime an LLM infers those values and a fixed policy allows or denies, making each decision auditable via the attributes and rule that fired. Evaluated on AgentDojo and AgentDyn with claimed strong IPI robustness. The auditability angle is the practical differentiator over pure LLM-judge authorization.

This mines 22,555 commits across 17,703 non-fork repos to study what happens when OpenAI, Anthropic, and Google retire models. The headline: an estimated 82% of migrations away from retired models happen after the shutdown date — after the app already failed. The share tracks notice length (89% post-shutdown for Anthropic's 60–114-day notices versus 13% for OpenAI's one-year Assistants notice). Model IDs are hard-coded in 94% of cases; provider-abstraction layers didn't help. Coder agreement was validated (kappa 0.89–0.95).

First systematic study of cost-inefficient behavior in coding agents, over 1,200 trajectories from Claude Code and Mini-SWE-Agent on SWE-bench Verified. Three recurring patterns — subsumed retrieval, similar-script generation, test re-execution — hit 79–98% of tasks and account for up to 22.75% of cost. Across 10k held-out trajectories, developer-designed skills cut cost by up to 41.73%, roughly double agent-synthesized skills, while structure-aware retrieval sometimes raised cost by 28%. The clear finding: human-authored, trace-agnostic skills beat auto-generated ones.

Tool observations dominate SWE-agent context. This work compresses older observations into soft tokens while keeping the agent's own turns and the last K observations in text (LOHA), paired with a distillation method (ACD) that trains latent reading without behavioral drift. On SWE-bench Verified, K=3 cuts per-call context 43–57% but resolve rates drop (12.1% vs 14.5%, 21.8% vs 27.5%); K=8 recovers some. Under a 32K limit, compression nearly doubles resolves (21.1% vs 11.1%). The compression-versus-accuracy tradeoff is explicit.

Anthropic's official prompting guidance for Claude Opus 5.5, covering the new model's behavior and how to steer it for tool use and agent workflows. It's a vendor doc, not engineering analysis, so treat it as the reference for what changed rather than an independent evaluation. Worth a scan if you're already on Opus and need to know which prompting patterns the new version rewards or breaks versus prior releases.

Imp ports DSPy — the declarative prompt-optimization framework — to the BEAM, published as an Elixir Hex package. For teams already running on Erlang/Elixir infrastructure, this means DSPy-style program compilation and optimization without shelling out to Python, with the BEAM's concurrency model underneath. It's a concrete framework artifact rather than a paper; the interesting question is how faithfully the optimizer semantics survive the port and whether the ecosystem of DSPy modules comes along.

Willison's WeAreDevelopers closing keynote is his usual chronological tour of the year in LLMs, delivered as annotated slides plus a YouTube recording. For this audience the useful threads are coding agents and eval practice, though the format is a broad survey rather than a deep look at any single mechanism. If you follow his blog you'll recognize most beats; the value is the consolidated timeline and his read on which trends actually stuck.