Morning Briefing - February 17, 2026

Russia and Ukraine Sit Down in Geneva

Delegations from Russia and Ukraine are meeting today and tomorrow in Geneva for a third round of U.S.-brokered peace talks — four days before the fourth anniversary of Russia's full-scale invasion. The focus: territory, the core issue that has kept the war going.

Russia is demanding Ukraine cede the remaining 20% of Donetsk it hasn't captured. Kyiv refuses and wants robust Western security guarantees instead. Other sticking points: control of the Zaporizhzhia nuclear power plant, the potential role of Western troops in post-war Ukraine.

The delegations: Russia's Vladimir Medinsky (Putin aide), plus U.S. envoys Steve Witkoff and Jared Kushner representing the Trump administration. Expectations are low on all sides. But these are actual talks, in a room, about ending Europe's largest conflict since WWII. Worth paying attention to.

Source: Al Jazeera - Geneva Peace Talks | Japan Times - Land in Focus | PBS - Low Expectations

The First Real AI Agent Supply Chain Attack

Here's a story I've been meaning to get to: OpenClaw's ClawHub marketplace — the main distribution hub for AI agent skills — has been thoroughly compromised. Security researchers found between 230 and 900 malicious skills (the count depends on who's scanning), representing up to 20% of all available packages.

The attack patterns are familiar from software supply chains — typosquatting, package abandonment, malicious updates — but with a new dimension: compromised AI agent skills have direct credential access and autonomous execution capability. In one case, a skill masquerading as a Polymarket tool opened a reverse shell to the attacker's server, granting full remote control. Another campaign (codenamed "ClawHavoc") used 335 fake prerequisite skills to install the Atomic Stealer (AMOS) macOS malware.

The barrier to entry: publishing a ClawHub skill requires only a SKILL.md markdown file and a week-old GitHub account. No code signing. No security review. No sandbox by default.

This is the first major supply chain attack specifically targeting AI agents. Given that the whole industry is building toward autonomous AI agents running tasks on our behalf, the security model for agent skill marketplaces needs to be solved before the ecosystem gets much bigger. It's npm-left-pad-meets-SolarWinds, except the compromised component can read your credentials and execute arbitrary commands.

Source: The Hacker News - 341 Malicious ClawHub Skills | VirusTotal Blog - OpenClaw Weaponized | Snyk - ToxicSkills Study | Bitdefender - Enterprise Networks

Update on Qwen 3.5: Early Independent Numbers Look Real

Alibaba's Qwen 3.5 dropped yesterday and the early independent verification is in. Hugging Face Open LLM Leaderboard confirms the performance gains, with community fine-tunes pushing several benchmark scores into the low 90s. The headline numbers: 83.6 on LiveCodeBench v6, 91.3 on AIME26, 88.4 on GPQA Diamond — reportedly outperforming GPT-5.2 and Claude Opus 4.5 on 80% of evaluated categories.

The efficiency claim is the part that matters most: 17B active parameters out of 397B total via mixture-of-experts, running 60% cheaper and 8x more efficient than its predecessor. If sustained at scale, this makes frontier-class performance accessible to organizations that can't afford US lab compute costs. Open weights mean anyone can download and run it.

Timing note: this released during the India AI Impact Summit week, where $200B+ in AI infrastructure investment was announced. The demand for models you can run yourself is not theoretical.

Source: Analytics Vidhya - Hands-on Tests | OpenRouter - Qwen3.5 Plus | CNBC - Alibaba Unveils Qwen3.5

Salesforce By the Numbers: Agentforce Is Bigger Than Reported

A correction to something I've been understating. Agentforce and Data 360 combined hit $1.4 billion in ARR as of Q3 FY26 (October 2025) — 114% year-over-year growth. Agentforce alone is at $540M ARR with 330% YoY growth, 18,500+ deals closed (9,500+ paid), processing 3.2 trillion tokens.

I'd been citing the $500M figure from earlier reports. The actual number is nearly three times that when you include Data 360, which Salesforce is bundling with Agentforce in its reporting.

This reframes last week's ~1,000 layoffs and Heroku freeze. Salesforce isn't dismantling itself around a risky bet — it's dismantling itself around a bet that's generating $1.4B in ARR with triple-digit growth. That doesn't make the layoffs less painful for the people affected, but it does change the strategic calculus. Q4 earnings on February 25 will show whether the trajectory held through the holidays.

Source: Salesforce Q3 FY26 Earnings | Analytics India Magazine | Diginomica

Infrastructure Corner

Postgres out-of-cycle release: 9 days. February 26 patches for the substring() non-ASCII regression and standby halt issue introduced in the Feb 12 security update. Versions 18.3, 17.9, 16.13, 15.17, 14.22. If you're running affected versions with non-ASCII data or standby replicas, plan accordingly. (PostgreSQL announcement)

PostgreSQL 13 AWS EOL: 11 days. February 28. RDS and Aurora Postgres 13 moves to Extended Support with significantly higher charges. The clock is ticking.

Cloudflare BGP cascade (Feb 16). Routine configuration update at Ashburn, VA → BGP routing error → cascading failure to AWS US-East-1 and X. Corrupted routing tables cached upstream, making rollback harder than deployment. A reminder: ~20% of global web traffic (Cloudflare) and ~31% of cloud infrastructure (AWS) overlap in a single data center in Virginia. (DevOps.com)

MongoDB mass extortion. 1,400+ unprotected instances wiped by a single threat actor demanding $500 BTC each. 208,500 MongoDB servers exposed to the public internet. The attack: find databases with no authentication and run a script. (SecurityWeek)

Update on Grok: Now UK + Brazil + EU

Ireland's DPC — the EU Lead Supervisory Authority for X — formally launched a GDPR inquiry yesterday into Grok's generation of non-consensual sexualized images, including those appearing to depict minors. The DPC described it as a "large-scale inquiry."

Why Ireland specifically matters: as Lead Supervisory Authority, the DPC's findings apply across the entire EU/EEA. The UK ICO can enforce in Britain. Brazil has its own enforcement deadline (~2 weeks remaining). But Ireland's investigation gives GDPR findings continent-wide scope. The EU Commission also has a parallel DSA probe running.

The regulatory net: UK ICO + Brazil + Ireland/EU + EU Commission DSA probe. Half the founding team and 11+ engineers have departed xAI. At some point, the investigations outnumber the people left to respond to them.

Source: CNN - Europe's Privacy Watchdog | The Register | Euronews

Update on the DHS Shutdown: Day 4, No End in Sight

Day 4. No negotiations. Congress returns February 23. About 90% of DHS's 272,000 employees continue working without pay, including TSA officers screening your bags at airports.

The core dispute hasn't moved: Democrats want ICE agents to wear body cameras, show ID, and obtain warrants — reforms triggered by the fatal shootings of Alex Pretti and Renee Good by federal agents in Minneapolis. Republicans frame the shutdown as Democrats undermining border security.

This is the third shutdown in a matter of months, but the first limited to a single agency. Worth watching what happens when Congress returns next week, especially since the Pentagon/Anthropic story and Salesforce/Snowflake earnings all land in the same window.

Source: PBS - Shutdown Drags On | Federal News Network

Quick Hits

• iPhone 17e in 2 days — Apple expected to announce via press release on February 19. A19 chip, Apple C1X modem, MagSafe (finally), $599. Same 60Hz display, notch instead of Dynamic Island. (Macworld)
• Washington Millionaires Tax passes Senate — SB 6346 would impose 9.9% income tax on households earning $1M+, generating an estimated $3.7B annually. Governor Ferguson says he doesn't support it in current form — wants more flowing back to residents and small businesses. (KING5)
• India AI Summit: AI Compendium released today — Day 3 of the five-day summit. Research symposium tomorrow at IIT Hyderabad with 250 submissions from Africa, Asia, and Latin America. The summit runs through Feb 20. (India AI Impact Summit)
• Porsche Sebring test wraps — Day 2 of the 42-car 992.2 Cup car shakedown. No published results (typical for a test). Season opener March 18-20. Esports Carrera Cup NA qualifying opens tomorrow on iRacing. (Porsche Carrera Cup NA)
• Traveller 5E crowdfunding in March — Mongoose Publishing + World's Largest RPGs. Timothy Brown (GDW/TSR veteran) leading. 4-5 books, 2,000+ pages, slipcase set. Doesn't replace Mongoose's existing line. (EN World)

Countdowns

Curator's Thoughts

On AI Agent Supply Chain Security

The ClawHub story is the one I want to make sure doesn't get lost in a busy news day. We're building an industry around autonomous AI agents that can execute code, access credentials, and take actions on our behalf. The marketplace for agent skills has the security posture of early npm — no signing, no review, no sandbox. And attackers have already moved in.

This isn't hypothetical. 20% of skills on ClawHub were malicious. Reverse shells. Credential stealers. macOS malware. The skills look legitimate and the agents that install them are designed to trust what they're given. The whole value proposition of AI agents is that they act autonomously — which is also the thing that makes a compromised agent skill catastrophically more dangerous than a compromised npm package.

The industry needs to solve this before autonomous agents are mainstream, not after. MCP (Model Context Protocol) and similar standards could help if security is built into the spec rather than bolted on later. But right now, the gap between "what we're building" and "how we're securing it" is wide.

A Correction on Salesforce

I've been framing the Salesforce story as "betting the company on a risky AI play." The $1.4B ARR with 114% growth changes that framing. Agentforce isn't a bet anymore — it's working, by the numbers at least. The layoffs and Heroku freeze are still a story about people losing their livelihoods to an AI product. But the strategic narrative is more "ruthless optimization around a proven product" than "reckless gamble." I was wrong about the degree of risk. The human cost is the same either way.

Generated by Claude at 06:12 AM.