Skip to content
Menu

RESEARCH / OPTIMIZATION LIMITS / OCTOBER 9, 2026

Check the plan.
Show the limits.

Imagine several systems passing work between them. Where would you place review points so every listed connection gets attention? Try a small fictional plan below, then compare it with the smallest set. OpenAI’s Unique Games research helps explain the limits of claiming the best plan in general.

Working browser example with fictional connections. No security product or account is connected.

The interactive example needs JavaScript. For a simple case, imagine four connections sharing one service: reviewing that service covers all four modeled connections. A general algorithm may select two endpoints instead. Both cover the connections, but only the one point set is smallest.

Why checking a plan matters.

Finding the best plan means comparing many possible choices. Checking a proposed plan means inspecting whether every listed connection has a review point. Those are different jobs. The example makes the coverage visible and gives an exact comparison for a tiny model.

Technical detail: what the Unique Games result says

OpenAI reports a Lean formalization of Unique Games hardness and related approximation thresholds. For general Vertex Cover, the reported result implies that a polynomial time algorithm with any fixed approximation factor below two would yield a polynomial time algorithm for 3SAT. Under P ≠ NP, this rules out that guarantee in the worst case. Read the reported scope.

This is a limit on general optimization guarantees. Small examples and special graph structures can still have exact solutions. It does not supply a faster detector, establish threat accuracy or make every security problem hard.

The model behind this example.

The model is called Vertex Cover. Review points are the circles, or vertices, and connections are the lines, or undirected edges. The task is to select as few points as possible while touching every edge. Real review costs, direction, missing records and whether endpoint review is adequate would need a separate model and evaluation.

A simple algorithm with a checkable bound.

The suggestion uses an established method: choose a link whose endpoints have not been selected, include both endpoints, and repeat. Those chosen links share no endpoints. Any complete cover needs at least one point per chosen link; selecting both uses at most twice the minimum. This method predates the linked research.

The page checks subsets in increasing size until it finds a complete cover for these tiny examples. This establishes an exact minimum for the model. The search grows exponentially and is capped at 12 points. The lower bound comes from the disjoint links, not from a prediction or a threat score.

What this could mean for your team.

Swarm Hunter: review a proposal.

A future planning feature could suggest which records or checkpoints to review, show every modeled connection, and separate complete coverage from claims of optimality. Analysts would still examine the evidence and missing context. Review Swarm Hunter’s current scope.

Agent Defender: preserve hard permissions.

A suggested plan cannot authorize an action. Permission checks must still use the approved task, resource and action boundaries. Checking concrete rules is a separate question from optimizing a review plan. Review Agent Defender’s current scope.

Research status and attribution.

The repository attributes the manuscripts to an internal OpenAI model and says verification stages vary. Its formalization catalogue lists the Unique Games main result and marks its overall review status “unchecked” as inspected October 9, 2026. Praxenor has not independently built or audited the full Lean proof.

This page is our experimental planning application. It does not establish product effectiveness, formal verification of Praxenor software or endorsement by OpenAI.