On Sunday, November 15, 2026, the United States and eight allied governments signed an expanded Pax Silica declaration in San Francisco aimed at jointly protecting AI-related semiconductors, critical minerals, energy, and industrial capacity from theft, coercive dependencies, and unauthorized transfer. If you run Claude Code, Codex, Cursor, or self-hosted agent stacks, you are not a signatory—but you inherit the supply-chain choices those governments are trying to harden: which frontier models, hosted inference APIs, MCP servers, and skills your employer still treats as trusted stack layers.
Poland Insight’s November 15, 2026 report, citing the Financial Times, names the ceremony’s nine signatories—the United States, Japan, Australia, Israel, the United Kingdom, South Korea, India, Qatar, and the United Arab Emirates—and says they will identify technologies and infrastructure requiring protection, coordinate supply-security measures, and develop AI-linked industrial and investment hubs through dedicated working groups. The same piece places Germany, Sweden, the Netherlands, and the European Commission as expected signers the following week, after harder EU negotiations than the bloc’s June 2026 Pax Silica accession.
explainx.ai has the full news timeline on Pax Silica’s joint supply-chain list for readers who want parallel coverage; this article stays on mechanism and defenses for coding-agent operators.
TL;DR
| Question | Short answer |
|---|---|
| Am I affected? | Indirectly—if your agents depend on cross-border models, MCP tools, cloud GPUs, or unaudited skills that procurement may reclassify |
| Is there a patch? | No product update; watch vendor lists, cloud regions, and version pins |
| Does data leave my machine? | Pax Silica does not mandate telemetry from your IDE; corporate policy may restrict which endpoints agents may call |
| What changed on Nov 15? | Operational declaration with joint sensitive-tech selection, diversification intel sharing, and anti-leakage coordination |
| What should I do today? | Inventory agent dependencies, scan new skills/MCP configs, document approved inference routes |
| Is this the same as export bans on chips? | Related but broader—covers software platforms, foundation models, and logistics, not only fabs |
What Pax Silica already committed to
The U.S. Department of State Pax Silica page publishes the foundational declaration language partners have been signing since the initiative launched. The text is explicit that economic value will flow across the global AI supply chain, including stacks that matter directly to coding agents:
- Software applications and platforms
- Frontier foundation models
- Compute and semiconductors
- Information connectivity, network infrastructure, and data centers
- Minerals refining, energy, and advanced manufacturing
The declaration also commits partners to cooperate on enforcement to protect sensitive technologies and critical infrastructure from undue access, influence, or control, and to address non-market practices that distort investment. That framing is macro-economic, but the listed layers are the same ones your agent touches when it calls an API, pulls a model weight, installs a skill, or attaches an MCP server.
Jacob Helberg, U.S. under secretary of state for economic affairs, is quoted across press relays—including Poland Insight’s summary of Financial Times reporting—as presenting the November 15 San Francisco Declaration as the next operational stage of Pax Silica, not a replacement for the original text on State.gov.
What the November 15, 2026 signing adds
Press relays agree on mechanics even though the full San Francisco Declaration text was not yet mirrored on State.gov when this post was drafted. According to Poland Insight:
- Nine countries signed in San Francisco on 15 November 2026 (listed above).
- Signatories intend to jointly select sensitive AI technologies and critical infrastructure for protection.
- Partners will share information on sectors needing supply-chain diversification.
- They will coordinate against technology leakage and unauthorized transfers that could give adversaries military or intelligence advantage.
- Working groups will examine supply-chain vulnerabilities and economic hubs meant to attract investment in protected industries.
- The package covers semiconductors, critical minerals, energy, and industrial capacity tied to AI data-center expansion.
Poland Insight also notes timing: the ceremony landed ahead of a planned meeting between U.S. President Donald Trump and Chinese President Xi Jinping, with rare-earth and mineral leverage still central to trade tension—relevant because GPU shortages and power constraints already shape which cloud regions agent hosts advertise as available.
Timeline: Pax Silica upgrades in November 2026
| Date | Event | Why coding-agent teams should notice |
|---|---|---|
| Jun 2026 | European Union joins foundational Pax Silica declaration (per Poland Insight / Commission press) | EU procurement law already intersects agent tooling for regulated firms |
| Nov 11, 2026 | EU member states authorize Commission to negotiate upgraded declaration text (Council document cited by Poland Insight) | Negotiation ≠ signature—expect policy churn while text stabilizes |
| Nov 12, 2026 | Euronews reports EU resistance to broad “sensitive technology” wording and superintelligence terminology | Legal definitions in the EU AI Act may diverge from U.S. “Super Intelligence (SI)” branding on State.gov |
| Nov 15, 2026 | Nine allies sign San Francisco Declaration (FT relay via Poland Insight) | Joint catalogs and working groups begin—details unpublished |
| Week of Nov 16, 2026 | Germany, Sweden, Netherlands, European Commission expected to sign (per same relay) | Watch for corporate guidance tying EU cloud regions to approved stacks |
Attack chain this is not (and what actually moves)
Pax Silica is not a prompt-injection demo or a malicious MCP rug pull. No signatory published a CVE against Claude Code on November 15. The risk pathway for practitioners is slower and organizational:
Policy signal (joint sensitive-tech list)
→ investment security & export-alignment reviews
→ procurement shrinks approved model/MCP/skill vendors
→ agents silently lose API routes or get replaced models
→ developers bypass policy with unaudited local skills (new supply-chain gap)
That last step is where local agent observability still matters: when approved vendors change, teams often import unaudited skills or self-hosted MCP wrappers to keep velocity. The F-Secure AI harms taxonomy for coding agents classifies supply-chain and tool-poisoning harms separately from geopolitical trade measures—both can land in the same sprint if policy moves faster than dependency review.
Semiconductor and minerals context (why agents feel it)
Poland Insight’s FT summary stresses spatial fragmentation: U.S. strength in chip design, Japanese materials and fab equipment, South Korean memory production, and China’s dominance in critical-mineral processing. Any disruption in those segments shows up for coding agents as GPU queue times, regional inference outages, or sudden model deprecations—not as a hook you can toggle off.
The piece also mentions industrial hubs combining semiconductor production, compute infrastructure, and energy. Locations and financing are unspecified, which means no reader should claim a datacenter map from press relays alone.
What this means if you run AI coding agents
Treat November 15 as a supply-chain governance signal, parallel to—but distinct from—the White House Super Intelligence Force push on AI incident reporting earlier in the week. SI Force discussed what to disclose after an agent incident; Pax Silica discusses which technologies partners agree are too sensitive to leak or over-concentrate.
Concrete checks grounded in published Pax Silica language and press relays:
- Model and API inventory — List every frontier or open-weight model your agents call, with region, vendor, and contract owner. The State declaration explicitly names foundation models and software platforms as strategic stack layers.
- MCP and skill provenance — Map MCP servers and skills to legal entities and hosting countries. Pax Silica’s anti-leakage coordination targets unauthorized transfer, not just malware; a benign skill hosted in a non-approved jurisdiction can still fail new procurement rules.
- Version pins before policy memos arrive — run
npm lsand keep lockfiles for agent-side packages; see npm supply-chain worms targeting AI tooling for why parallel dependency audits matter when policy shrinks vendor lists. - Energy and cloud-region assumptions — If your org routes agents through specific inference clusters for cost or latency, note power and mineral supply narratives in Pax Silica—outages may be explained as infrastructure, not application bugs.
- EU divergence watch — If you ship agents for EU customers, track whether Commission signature adds AI Act-aligned definitions that conflict with U.S. superintelligence framing reported by Euronews—contract language may change before your
mcp.jsondoes.
Commands and habits from Beam docs (not Pax Silica enforcement):
beam scan --mcp
beam scan path/to/skill
Those scans do not read export-control lists; they apply heuristic patterns (for example risky shell pipelines or missing MCP version pins) described in the MCP security practical guide.
How Beam handles overlapping supply-chain risk
Beam is local, on-device observability and heuristic scanning for coding agents—not an investment-security or export-compliance platform.
Where overlap is real:
- Skill and MCP config scanning before agents act on new instructions, with 11 heuristic patterns plus an MCP version-pin check (see skill testing cookbooks).
- Skill inventory on enrolled machines (
beam skills index/ sync to the workspace dashboard) and optionalbeam mcp proxyto redact secrets in MCP server responses before they reach the agent. - Hook-based enforce on supported harnesses (Claude Code and Codex PreToolUse) to deny known-risky tool shapes—complementary to procurement allowlists, not a substitute.
Plain limits:
- Beam does not implement Pax Silica catalogs, CISA-style sector rules, or foreign-investment reviews.
- Scanning is heuristic, not malware-grade or export-law-grade classification.
- Missing cached policy is fail-open on the collector; corporate MDM policy still wins for approved vendor lists.
For architecture that pushes supply-chain enforcement outside the agent entirely, compare Agent Baseline’s secure software factory reference design—registry and egress proxies the agent cannot reconfigure—with Beam’s per-machine watch layer on what we monitor locally.
What is verified vs. still missing
| Claim | Status | Source |
|---|---|---|
| Nine countries signed in San Francisco on 15 Nov 2026 | Reported (FT relay) | Poland Insight |
| Joint sensitive-tech lists and working groups are next steps | Reported intent | Same relay |
| Germany, Sweden, Netherlands, EU Commission signing following week | Expected, not confirmed at publish time | Same relay |
| Published text of the San Francisco Declaration on State.gov | Not found on the Pax Silica landing page at draft time | state.gov/pax-silica shows foundational declaration only |
| Enumerated technologies on the joint list | Unpublished | Press relays describe intent only |
| Penalties in partner jurisdictions | Unspecified | Poland Insight |
If your security team asks for primary .gov text, point them to State’s Pax Silica hub and watch for an updated press release; do not treat HuggingNews or social reposts as authoritative versions.
Frequently asked questions
The front matter lists eight schema-ready pairs; this section mirrors them for readers skimming the page.
What did the US and allies sign on November 15, 2026?
See the first FAQ in front matter: a San Francisco ceremony expanding Pax Silica with operational supply-chain commitments, per Financial Times reporting summarized by Poland Insight.
Which countries are Pax Silica members after the November 15 signing?
Distinguish two lists: the long-running Pax Silica signatory roll on State.gov (Argentina through the United Kingdom, plus EU accession in June 2026) and the nine-country core reported at the November 15 ceremony. Press searches mixing “Pax Silica Philippines” with this U.S.-led diplomatic initiative often describe unrelated local projects—verify state.gov and FT-dated articles before citing member lists.
Does this change whether Pax Silica is “good” for developers?
Geopolitical accords are not product reviews. For coding agents, the relevant question is whether your employer’s approved stack narrows after joint sensitive-technology selection—not whether the diplomacy is favorable in abstract.
Related reading
- White House SI Force: AI incident reporting and liability relief
- Agent Baseline secure software factory: enforce limits outside the agent
- MCP security practical guide: rug pulls and version pins
- F-Secure AI harms taxonomy for coding agents
- Recorded Future: AI incidents involving developer tools
- November 2026 attacks timeline
- Beam landing: on-device watches
Pax Silica signatory lists, EU negotiation status, and San Francisco Declaration publication on State.gov are accurate as of November 15, 2026. Beam is a local prototype; confirm hook coverage and scan heuristics in current docs before relying on enforce mode.

