A “minor update” that moves every row of its own table
DeepSeek published DeepSeek-V4.1-Flash-1017 to Hugging Face on Saturday 17 October, under an MIT licence and ungated. The card calls it “a minor update” to the V4.1-Flash whose repository dates from 10 September. All eleven rows of its comparison table improve, the largest by 11.1 points on Terminal-Bench 4.0 and 8.8 on ProgramBench. DeepSWE v1.1 goes from 74.2 to 77.0, the highest figure in that row, above Opus 5 at 74.0. Opus 5 leads the other three coding rows: 51.8 on Terminal-Bench 4.0, 37.0 on ProgramBench against 29.1, 75.3 on NL2Repo-Bench against 68.2.
Every figure there is DeepSeek’s, and so is the harness: the card’s footnote names Claude Code for Terminal-Bench 4.0 and Mini-SWE-V2 for DeepSWE v1.1. Artificial Analysis’s changelog carries no entry after 16 October, so the checkpoint has no independent Intelligence Index score. The release ships without a Jinja chat template. The card points to a Python reference implementation in an encoding folder, and to deepseek-recipe, “a set of Rust libraries with Python bindings”, with inference, tool execution and HTTP transport “left to the caller”. The weights need no approval and carry no revenue threshold.
Three authorisation fixes sit behind Google’s agent mesh, and two got CVE IDs this morning
NVD published CVE-2026-102503 and CVE-2026-103091 at 03:16 UTC today, both scored 9.3 critical by Google’s own CVE coordination team, with NVD’s own status on each still “Received”. They cover google/sam, the Go project now at github.com/google/agentmesh and named in both records as Google OSS Agent Mesh: a libp2p network carrying MCP tool calls between machines under Biscuit capability tokens. In the first, the MCP stream handler enforced WithBiscuitAuth and Authorize checks while a parallel /libp2p-http ingress routed requests to backend handlers without verifying a Biscuit token or mesh enrollment.
The second is the shipped default policy. GET /info disclosed the OIDC configuration, POST /register auto-enrolled any valid OIDC token holder without claim filtering or ban checks, and the default policy bound sam:system:authenticated to wildcard services and targets. The CVE text puts the result as full read, write and execution access on backend services. The advisories went up on 4 and 6 September, six weeks before the CVE IDs reached NVD; their fixes are dated 6 August and 7 September. A third advisory of that vintage, rated high, still has no CVE ID. Version 0.1.0 followed on 16 October, called “the first stable release” in its own notes; the README calls the project “not an officially supported Google product”.
Quick Hits
- Beam’s weights are out and score 30. Reflection published the 501B Apache-2.0 weights to Hugging Face on 16 October, eleven days after announcing the model. Artificial Analysis ranks it second among US open-weights models and tenth among all open-weights models.
- The Claude Agent SDK for Python got its CVE three months after the fix. CVE-2026-96620, 9.2 from GitHub, says an unvalidated resume value starting with a dash could inject --mcp-config into the CLI. Fixed in 0.2.121 on 17 July.
- Two AI consoles leaked credentials, one unauthenticated. CVE-2026-111704 (9.3, VulnCheck’s) says LangChat registered no Sa-Token interceptor, so /api/v1/auth/users returned plaintext passwords. CVE-2026-111682 (7.1, also VulnCheck’s) says any ROLE_USER account in EFAK 5.0.0 through 5.1.0 could read unmasked provider keys.
- A desktop coding agent connected to MCP servers declared by the repository it opened. CVE-2026-88755, 7.8 from CISA-ADP with NVD’s analysis deferred, says Z.ai ZCode v3.11.2.6792 does this automatically, so opening a crafted repository could run code.