GitHub AI Weekly · Issue 07 (Sep 26, 2026) · AI Beyond

Top 10 projects

1. 01

   cloudflare/security-audit-skill

   AI codingJavaScript

   This week+11,262
   

   A coding-agent skill for multi-phase security audits with independently verified, machine-readable findings

   Stars

   21,642

   Forks

   1,246

   Open issues

   50

   Last commit

   2026-09-14

   It leads this snapshot with 11,262 weekly stars. Cloudflare structures auditing into six phases, from reconnaissance and coverage-led hunting to independent verification and structured reporting. No new default-branch commits were found this week, so rising interest should not be described as a new feature release.

   Teams with owned or explicitly authorized code, security-review experience and isolated test environments can try it as a lead-generation and verification aid on a clearly bounded module.

   This is not Cloudflare's full fleet-wide vulnerability harness, and one run cannot guarantee complete coverage. Executing target code requires OS-enforced isolation, disabled external networking and resource limits; without those controls, leads should remain unvalidated.

   Deep dives

   Its useful distinction is not simply running more agents: the finder and verifier are separate. A fresh verifier attempts to disprove a candidate, and final records undergo another source check. That gives a conclusion an adversarial review instead of letting the same model rephrase its own suspicion as a confirmed report.

   The coverage ledger can be more useful than a polished findings list. It records checked surfaces, missing evidence and findings that need revisiting after code changes. The repository separates confirmed, needs_validation and rejected records, withholding severity from unresolved leads. This helps avoid overstating an untested result, but a complete ledger is still not proof that nothing was missed.

   We would evaluate it on internal samples with existing human findings, recording false positives, omissions and repeated-run costs. The README says a single run found roughly half the vulnerabilities found across repeated runs in the authors' tests; that is their observation, not a recall guarantee for arbitrary repositories. Without a security sandbox, begin with source review and evidence organization rather than broader execution privileges.

2. 02

   alibaba/open-code-review

   AI codingGo

   This week+6,920
   

   Secure, fast, efficient, battle-tested at Alibaba's scale. Hybrid architecture code review tool: deterministic pipelines + LLM Agent, precise line-level comments, built-in multi-language ruleset (NPE, thread-safety, XSS, SQL injection), OpenAI & Anthropic compatible.

   Stars

   41,322

   Forks

   2,970

   Open issues

   224

   Last commit

   2026-09-24

   Following last week's appearance, v1.12.8 and v1.12.9 shipped on September 21–22 with fixes for dropped quoted paths and findings lost after file renames. September 24 main-branch changes added a tool-call round-trip test and further non-ASCII path handling.

   Teams reviewing non-ASCII filenames, bulk renames or large PRs can use those edge cases for regression tests before adding it to routine review.

   These changes improve the review pipeline, not the model's ability to catch every defect. Check released and main-branch behavior separately, along with code sent to external models, credential storage and review budgets.

3. 03

   stablyai/orca

   AI agentsTypeScript

   This week+6,547
   

   Orca is the ADE for working with a fleet of parallel agents. Run any coding agent with your own subscription. Available on desktop, mobile and remote runtime.

   Stars

   78,284

   Forks

   5,121

   Open issues

   6,876

   Last commit

   2026-09-25

   Orca reached v1.4.212 this week, fixing Codex 0.157+ startup in Orca-managed homes. The preceding v1.4.211 improved native-chat tool state, child-task working state and remote-terminal reliability, shifting attention from parallelism to session continuity.

   Developers already using multiple coding agents across local and remote environments can first test the full start, disconnect, resume and stop lifecycle.

   Separate workspaces do not automatically eliminate shared-branch, port or external-service conflicts. Model subscriptions and runtime environments remain your responsibility. Verify child processes actually stop, and do not assume post-release fixes are in the installed package.

4. 04

   affaan-m/ECC

   AI codingJavaScript

   This week+6,193
   

   The agent harness performance optimization system. Skills, instincts, memory, security, and research-first development for Claude Code, Codex, Opencode, Cursor and beyond.

   Stars

   267,480

   Forks

   39,954

   Open issues

   234

   Last commit

   2026-09-24

   This week's main-branch fixes tightened execution boundaries: GateGuard detects destructive SQL quoted inside client commands, failed steps no longer supply successful evidence downstream, and the slide export server is confined to the deck directory and loopback.

   Existing ECC users can check which enabled hooks and workflows these fixes affect, then upgrade and regression-test those modules rather than installing everything.

   The latest release record remains v2.2.1, so this week's main-branch work may not yet be distributed through your installation channel. Command filtering does not replace an OS sandbox, least privilege, database backups or approval for dangerous operations.

5. 05

   Tencent/WeKnora

   RAGGo

   This week+3,749
   

   Open-source LLM knowledge platform: turn raw documents into a queryable RAG, an autonomous reasoning agent, and a self-maintaining Wiki.

   Stars

   30,067

   Forks

   4,035

   Open issues

   500

   Last commit

   2026-09-25

   The September 24 v0.8.2 release adds local-browser operation, workspace-scoped built-in MCP servers, and steering, forking and rewinding conversations. It extends knowledge Q&A toward tasks carried out using that knowledge.

   Teams with internal document collections can start with read-only retrieval, then selectively enable browser and tool permissions for controlled multi-step work.

   Upgrade notes make DingTalk channels Stream-only and say sandbox commands run as root; review container mounts and network boundaries. The README recommends private-network deployment. A login screen alone does not make direct public exposure appropriate.

6. 06

   vectorize-io/hindsight

   AI agentsPython

   This week+3,363
   

   Hindsight: Agent Memory That Learns

   Stars

   29,763

   Forks

   3,144

   Open issues

   126

   Last commit

   2026-09-25

   Hindsight enters the top ten after the September 21 v0.10.1 release, including attachment-handling and memory-defense fixes. It separates cross-session memory into retention, retrieval and reflection rather than merely replaying chat history.

   Agent teams repeatedly using preferences, project experience or support context can select one narrow scenario and compare it with a version without persistent memory.

   Fact extraction and reflection still depend on models and can accumulate errors, latency and cost. Secret and PII scanning is opt-in per bank; isolation, authentication, deletion and retention require your own acceptance tests.

   Deep dives

   Retain extracts facts, time and entities; recall combines semantic, keyword, graph and temporal retrieval; reflect reasons over the stored material. Those stages let a team distinguish incorrect storage, incorrect retrieval and incorrect inference. For long-running work, that is more actionable than saying the context window was too short.

   Knowledge pages are another useful idea: a standing answer is refreshed in the background and can be read without another model call. Cheap reads depend on background consolidation, however; they do not make the whole memory system free. Tests should cover newly corrected preferences, stale evidence and irrelevant experience from another project.

   We care about whether memory can be corrected, not only whether an old answer can be recovered. A pilot can include a subsequently withdrawn fact, users with identical names and a test secret, then check isolation, updates and removal. Published memory benchmarks can guide test selection, but cannot replace your dataset or establish universal superiority over ordinary RAG.

7. 07

   addyosmani/agent-skills

   AI codingJavaScript

   This week+3,345
   

   Production-grade engineering skills for AI coding agents.

   Stars

   99,070

   Forks

   10,403

   Open issues

   124

   Last commit

   2026-09-23

   This week's main-branch work examines the skills themselves: it restores security rules lost during text condensation, validates links inside reference files, applies YAML validation to command frontmatter, and pilots code-review plugin evaluations.

   Maintainers of team skill libraries can borrow its reference, trigger and evaluation checks, particularly after shortening instructions.

   The latest formal release remains 0.6.10; this week's validation and evaluation changes should not be assumed to ship in it. Valid formatting does not prove effective instructions, and host-specific triggers, permissions and hooks still need testing.

8. 08

   bojieli/ai-agent-book

   AI agentsPython

   This week+2,481
   

   《深入理解 AI Agent：设计原理与工程实践》（李博杰 著）开源主仓库：全书正文、编译版 PDF 与按章配套代码

   Stars

   50,994

   Forks

   5,718

   Open issues

   10

   Last commit

   2026-09-25

   On September 23, AI Agents in Depth distilled its material into 22 Claude Code skills and added guidance on context-prefix constraints for preserved thinking, with synchronized translations. It connects a textbook with methods usable during engineering work.

   Readers comfortable with Python, Git and APIs can build a small chapter-driven project and use failure cases to test their design decisions.

   The author explicitly says the companion code is generated by coding agents; it should not be treated as line-by-line human-audited production code. Latest ebook assets are overwritten in place, so record a commit for reproducibility. Readers manage their own model costs and credentials.

9. 09

   anthropics/claude-code

   AI codingTypeScript

   This week+2,384
   

   Claude Code is an agentic coding tool that lives in your terminal, understands your codebase, and helps you code faster by executing routine tasks, explaining complex code, and handling git workflows - all through natural language commands.

   Stars

   148,097

   Forks

   24,573

   Open issues

   13,055

   Last commit

   2026-09-25

   At this snapshot Claude Code has released v2.1.283, adding prompt-audit for older prompting patterns, stricter model allow/deny settings, and fixes for background MCP progress and server cleanup when sessions end.

   Development organizations with accumulated instruction files and plugins can start by reviewing old prompts and defining permitted model versions, rather than focusing only on terminal UI changes.

   A public repository does not mean the entire product is an open-source implementation you can self-host. Assess whether tool-content telemetry could capture code or business data before enabling it. Prompt-audit is an aid, not proof that a prompt revision is better.

10. 10

    anthropics/financial-services

    AI agentsPython

    This week+2,382
    

    Stars

    37,549

    Forks

    5,432

    Open issues

    222

    Last commit

    2026-09-21

    Anthropic's financial-workflow repository enters this week's top ten with reference agents and connectors for research, modeling and reconciliation. The only default-branch commit found this week removes the financial-advisors directory; the broader collection should not be presented as newly launched this week.

    Teams with financial expertise, licensed data and qualified reviewers can test research-note or reconciliation drafts on sanitized samples, measuring traceability and review effort.

    The README requires professional sign-off: these templates do not execute trades or post to ledgers and do not constitute investment advice. Data access may require separate subscriptions. The README still lists the deleted directory, and subagent delegation remains a research preview.

    Deep dives

    The repository separates skills, commands, connectors and end-to-end agents. The same prompts and skills can be used through a plugin or a managed-agent wrapper. Its value is less about naming an agent after an analyst than specifying inputs, deliverables and the point where a human must sign off.

    For a financial team, a plausible document is only the beginning. Data providers, reporting periods, traceable spreadsheet formulas and approval for external delivery matter more than extra conclusions. We would compare a previously completed human task, preserving source data and revisions. A working connector is not evidence that the organization has licensed the data.

    A directory deletion with a lingering README entry is a practical warning: star growth does not guarantee installation instructions are current. Pin the repository revision, check the actual plugin inventory and inspect permissions before adoption. Templates cannot substitute for your approvals around external communication, transactions, ledger posting or client onboarding.