Mistral's Le Chonk weights drop; GitLab flaw exploited

Also: Azure retires a product with no successor, and three upgrades worth pasting today.

The main thing

If your team wants a self-hosted model that can write code, your shortlist just got one trillion parameters heavier. Read the license before you get attached.

Mistral has released the weights of Mistral Large 4, which it nicknames Le Chonk, so nobody can accuse us of being rude. The model is multimodal, trained from scratch on European infrastructure, and has had three more weeks of reinforcement learning since its public preview. It also gains three reasoning modes for trading performance against cost.

These are Mistral's reported results, so treat them as the vendor's case:

• Artificial Analysis Coding Agent Index: 48%, effectively tied with GLM-5.3, using 47% fewer tokens per task
• Terminal-Bench 4.0: 41%, up 13 percentage points on the preview, with 30% fewer output tokens
• FrontierCode 2.0 (Cognition's private benchmark): 19.7%, which Mistral says leads
• Artificial Analysis Cyber Index: 51%
• Malicious cyber requests: 96.6% average refusal rate

The security pitch comes with a caveat. Mistral says it sees "a potential for uplift to conduct real attacks" and is changing its guardrails.

The Hacker News discussion filled some gaps. One commenter says the license is a modified MIT that isn't OSI-approved, and that serving the model as a business requires a license. Another reports hosted pricing fell from $1.36/$4.18 to $0.7/$2.1 per million input/output tokens, enough to warrant "a major reassessment."

What to do today: If self-hosting is on your roadmap, grab the weights from Hugging Face, confirm the license and hardware needs yourself, and run your own evals. Otherwise, nothing.

Quick bits

• GitLab's CVE-2026-97229 is on CISA's exploited list. On CE and EE, an unauthenticated user could execute code through a crafted regex in CI/CD configuration. Our entry names no fixed versions, so self-hosted admins should check GitLab's advisory now.
• Azure Deployment Environments retires February 22, 2027 with no successor; Microsoft says to use Bicep or ARM directly. Dev Box follows September 18, 2028, pointed at Windows 365, which Microsoft says isn't one-to-one. Start your inventory.
• Claude Code v2.1.314 adds permissions.blockWritesOutsideWorkingDirectories. Write, Edit and NotebookEdit refuse paths outside the working directories in every permission mode; shell commands aren't on that list. Also, a failed claude -p startup now exits 1.
• Cisco ASA and FTD flaw CVE-2026-76418 is being exploited, per CISA. An unauthenticated remote attacker can force a device reload, causing denial of service. Our entry names no fixed versions, so network teams should check with Cisco.
• Keycloak authorization bypass, CVE-2026-9800. Any authenticated user can skip policy checks by putting the access-denied path in a URL. Upgrade to org.keycloak:keycloak-services:26.6.4 and org.keycloak:keycloak-policy-enforcer:26.0.10; earlier versions are affected.
• GitHub added webhook IP ranges: 9.234.98.160/28, 9.234.98.176/28 and 9.234.106.48/28. Add them to any static allowlist or risk rejected deliveries. Syncing from the Meta API? Nothing to do.
• Google's OAuth platform now supports DPoP, so a stolen token can't be replayed without the client's private key. It covers refresh tokens only. For multi-agent delegation, Google points developers to its Secure Token Service.
• Attackers chained PHP2Shell with a kernel exploit to root 25 PAN-OS servers, Dream reports. The chain uses CVE-2026-109618, CVE-2026-109619 and CVE-2026-46300. PAN-OS admins should apply Palo Alto's fixed versions and restrict management-interface exposure.

Also happening

• Linux TCP use-after-free, CVE-2026-98432: Linux CNA scores it 9.8; install your vendor's fixed kernel and reboot, MPTCP listeners first.
• Android zero-click SMS crash: one WAP Push message kills com.android.phone; Google's fix covers Android 14 through 17 only.
• Twisted HTTP/2 Rapid Reset DoS: versions through 26.4.0 are affected; first patched version is Twisted==26.10.0rc1.
• shopify_app cross-shop bypass, CVE-2026-63414: 22.1.0 up to, not including, 23.0.3, when token-exchange apps scope by current_shopify_domain or params[:shop]; upgrade to shopify_app:23.0.3.
• cert-manager Challenge bypass, CVE-2026-62290: namespace users can misuse ClusterIssuer DNS credentials; upgrade to github.com/cert-manager/cert-manager@1.19.6 or github.com/cert-manager/cert-manager@1.20.3.
• PraisonAI sandbox escape, CVE-2026-61438: YAML script: steps can run OS commands when PRAISONAI_ALLOW_JOB_WORKFLOWS=true; upgrade to praisonai==4.6.78.
• Binance suspends withdrawals on multiple networks: loss estimates start at $2 billion, unconfirmed; Binance says user funds are safe.
• AO3 offline for two to four weeks: database theft can't be ruled out, so change any reused password.
• Cloudflare's Q3 DDoS report: about 8,226 attacks per hour on its network, 96% over within 10 minutes.
• ChatGPT's $100, $200 and $500 plans gain API credits: the same quantity as the plan price, matching Anthropic's deal.
• Vercel AI SDK stabilizes two APIs: filterActiveTools and LanguageModelStreamPart graduate; experimental names are deprecated.
• Bug0 QA Agent: plain-English Playwright steps with cached replays, source-available under FSL.
• How the Claude Code team reviews code: Claude writes about 90% of it, and nobody keeps a complete mental model.
• Vercel Flags cold starts drop to 10 ms: p75 in internal testing; needs @flags-sdk/vercel@1.5.3 or later.
• AWS sales grew 39% as Amazon missed expectations: 2026 capex stays at about $220 billion, largely for AI.

Patch roundup

• decode-uri-component: decode-uri-component@0.5.1
• zip: zip@9.0.0
• @sylphx/pdf-reader-mcp: @sylphx/pdf-reader-mcp@3.0.1

Go patch something, then ask your GPUs how they feel about the name Le Chonk.

← All issues