Scan window: 2026-09-14 16:02:27Z to 2026-09-15 16:00:31Z; retrieval overlap starts 13:02:27Z. Four new events, one update, two duplicate builds skipped.
Daily Executive Summary
- Four research signals focus on agent safety verification, skill routing, and repository-scale evaluation. Plan injection can steer actions while leaving a plausible reasoning trace; VLoc Bench finds that security agents still struggle to locate flaws and abstain after fixes.
- Gavel explores routing from a model’s internal states, while AlgoEvo uses execution feedback for algorithm search. Neither paper links a public implementation yet, so both remain watch items. Claude Code 2.1.271 adds per-command network authorization and fixes permission checks.
Models
- No in-window foundation-model or GA API release met the admission bar.
Agent & AI Engineering
- Plan injection (TRIAL): In controlled tests, actors restate injected plans as their own reasoning and monitors miss 25–33% of attacks. Test trace monitoring separately from action authorization.
- Gavel (WATCH): Frozen-model internal states route skills with gains up to 13.4 points on written tasks. Await code and held-out catalog tests.
- AlgoEvo (WATCH): Execution feedback changes the search path. Authors report stronger results at lower budgets on six tasks; exact deltas and code are not available.
Open Source
- None of the four arXiv abstract pages links a directly reproducible implementation repository. No new stable open-source tool release met the bar.
Research
- VLoc Bench (TRIAL): Its 500 real vulnerabilities span 290 repositories. The best File F1 is 0.229; no evaluated model locates the flaw in 38.4% of tasks, and post-fix abstention is also weak.
- Plan injection and the earlier OATS study show that clean text or skill configuration does not establish action safety. This extends trend #7 rather than creating a new trend.
Developer Tools
- Claude Code 2.1.271 (material update): Sandboxed Bash, PowerShell and Monitor get per-command allowed_domains. Multiple organization-policy, managed-MCP and Bash permission gaps are fixed. 2.1.272 documents bug fixes only.
- Codex 0.155 remains alpha. The September 15 Gemini CLI nightly uses the same source hash as prior builds and is skipped as a duplicate. LiteLLM 1.101 stable does not promote the 1.102 RC.
Infrastructure
- No in-window serving or chip-infrastructure release met the bar.
Business & Policy
- No verified event changed model supply, API cost, chip access or the open-source landscape.
Trend Signals
- Trend #7 (emerging / Medium): Plan-injection tests and Claude Code per-command authorization add evidence; a second platform’s GA consequence policy and adoption telemetry are still missing.
- Trend #3 (established / High): VLoc Bench measures security-agent capability but does not change the multi-agent runtime adoption assessment.
- No new trend has enough independent evidence in this cycle.
Tech Radar
- New plan-injection monitoring test (Agent Security / TRIAL): plausible traces can hide malicious plans, so action-level acceptance testing is warranted
- New Gavel (Agent & AI Engineering / WATCH): internal-state routing may save catalog context, but code is not public
- New VLoc Bench (Agent Evaluation / TRIAL): test localization and post-fix abstention separately
- New AlgoEvo (Agent & AI Engineering / WATCH): feedback-guided search awaits budget-matched reproduction
- Claude Code (Developer Tools / ADOPT): per-command network policy and permission fixes reduce sandbox exposure; review upgrade settings
- Current radar: ADOPT 6 / TRIAL 67 / WATCH 84. The site insights page holds the full current radar.
Worth Trying
- Add an externally planted plan to an agent test, then measure both trace monitoring and action authorization.
- Use VLoc Bench’s vulnerable/patched pairing to test security-agent file localization and abstention.
Watch Items
- Await public implementations and independent reproduction for Gavel, AlgoEvo and VLoc Bench.
- Watch for a second agent platform with unified shell, MCP and browser action policy plus adoption telemetry.