CBAI fellowship and the test-time distribution-shift research front — formal verification becomes the answer to the test-environment problem
The CBAI Summer Fellowship's formal-verification track is the methodological response to a deep alignment problem: when models distinguish test from deployment, proving safety properties matters more than measuring them.
The CBAI Summer Fellowship opens with explicit formal-verification and multi-agent-safety tracks — both responses to methodological challenges that emerged in the past 12 months.
The methodological pivot
Formal verification proves safety properties under stated assumptions. The proof doesn't depend on the model failing to detect the test context — it's a mathematical claim about behavior across input spaces. For the test-environment distinction problem, formal verification is structurally a better fit than red-teaming.
Why multi-agent safety matters now
Multi-agent safety covers what happens when Claude Code agents orchestrate Devin agents, when Codex agents call out to MCP tools backed by other agents, when production deployments compose multiple autonomous agents. The safety properties of any individual agent don't compose into the safety properties of the agent-network. That's the research front that needs new methodology — and the CBAI track expansion is the field acknowledging it.
The pipeline math
CBAI 2026 + MATS Summer 2026 + Anthropic Claude Corps = three pipelines feeding the AI-safety-trained labor market entering 2027. Combined cohort size is substantially larger than any prior year. The headcount constraint on alignment work — historically the binding factor — is structurally easing. For frontier labs trying to staff safety-engineering teams in 2027, the labor market will be measurably better than 2024-2025.
The synthesis-paper anchor
"An Approach to Technical AGI Safety and Security" is functioning as the curriculum-reference synthesis paper for the 2027 cohort. That's a sign of field maturity — a synthesis paper anchored to specific methodological pivots, rather than a contested-claim paper that splits the field.
Opportunity Desk — CBAI Summer Research Fellowship in AI Safety 2026 → · ArXiv — An Approach to Technical AGI Safety and Security →