South Park Commons trained an overlay that means 'do not execute' and put it over untrusted text in five benchmarks on Qwen 3.5 9B, achieving 99.0% SEP separation, 6.2% TensorTrust hijacking, 99% AlpacaFarm attack success, 0.1% Quadrat-IPI compliance, and 0% PIArena success.
Public source
Publisher name
Public post
I seem to have fixed prompt injection? Every prompt injection attack works the same way: executable instructions get planted in text that was never meant to be executed.…
Company
South Park Commons
You only get a few shots at building your life’s work.
- Industry
- Venture Capital and Private Equity Principals
- Location
- San Francisco, US
- Company size
- 201–500 employees
About South Park Commons
South Park Commons is a community of technologists and builders dedicated to helping each other explore, learn, and figure out what's next—or as we call it, go from -1 to 0. We believe great ideas very rarely come about in isolation. The SPC Fund invests in companies that emerge from our members' explorations, ensuring the community's longevity and supporting the next stage of our founders' journeys.
See moreLatest activity
Latest activity from South Park Commons
15 signals
Presence & Recognition
South Park Commons is hosting Rob Moyer's Co-Sell Summit in San Francisco this week for partnership minds.
People
South Park Commons announced that Qasar Younis, co-founder and CEO of Applied Intuition, will join the company on September 15th.
Presence & Recognition
South Park Commons hosted a discussion on building production-grade AI systems.
Discover more
Similar signals
Similar public activity from other companies.
Research & Knowledge
Enclave
Enclave ran 105 runs across seven frontier models, with GPT 5.6 Sol and GLM 5.3 solving both Jenkins challenges.
Research & Knowledge
Ridge Security Technology Inc.
Ridge Security Technology Inc. published the industry's first public benchmark comparing leading LLMs in autonomous penetration-testing workflows.
Research & Knowledge
Semgrep
Semgrep benchmarked OpenAI's new Astra model against 13 points higher than Luna, scoring 100% on ExploitBench.
Research & Knowledge
McAfee
McAfee published a comparison of 8 abliterated Qwen 27b 3.8 models using 167 GPU core hours, marking the most capable dense ~30b LLM to date.
Research & Knowledge
BINARLY