Write bot-resistant code with KillSwitch, a programming language designed to confuse LLMs

Cloudflare engineer Dominik Picheta created KillSwitch, an open-source, deliberately obfuscated programming language intended to test and confuse LLMs. The best-performing model scored 66.8% on KillSwitch coding tasks, while most scored below 50%; Picheta said the benchmark tasks were simple and expects models to struggle with more complex scripts. The article describes an experiment and language release, not evidence of a security breach or a material company development.

Analysis

This is evidence about evaluation fragility, not a demonstrated security product or commercial initiative at Cloudflare. The key market implication is that coding-agent performance on familiar, benchmark-shaped tasks may overstate reliability on unfamiliar inputs. If that gap persists, enterprises may capture less near-term labor-cost reduction than AI-productivity narratives imply, while spending more on code review, testing, and model red-teaming. That is a conditional sector-level mechanism, not a revenue signal for NET.

The benchmark is developer-created, apparently small, and not independently replicated; its score should not be treated as a general measure of model capability. Nor is an open-source obfuscation technique a durable moat: exposure through scraping or targeted training can erode the test’s novelty. The more durable issue is evaluation design—models can score well on visible task patterns yet fail on distribution shifts.

Near term, no material earnings or valuation catalyst for NET is established. Over 1–3 months, watch for independent replication on unseen, more complex code and for enterprise evidence on coding-agent error rates and review costs. Over 6–18 months, persistent reliability gaps could temper software-labor productivity assumptions and support demand for testing and security validation; rapid improvement on held-out tasks would weaken that case. The contrarian point is that this is more a warning against overinterpreting benchmark wins than proof that AI coding adoption is stalled.

Key Decisions for Investors

• No trade in NET on this item: an employee’s open-source experiment does not establish a Cloudflare product, customer demand, or consolidated financial exposure. Reassess only if Cloudflare formally commercializes related tooling or discloses measurable adoption.
• Treat coding-agent productivity claims as an evidence question, not a blanket short. Monitor independent tests on previously unseen, complex code, plus enterprise defect rates, review hours, and paid-seat adoption over the next 1–3 months.
• Watch software testing, code-security, and AI red-team providers as potential beneficiaries if validation workloads rise; do not initiate a sector position without evidence that incremental spend is converting into orders or guidance.
• Falsify the cautious-productivity thesis if independent held-out evaluations show sustained improvement alongside lower enterprise review costs and rising coding-agent usage; strengthen it if error rates and human-review burdens remain high despite model upgrades.