OpenAI Publishes Wet-Lab Benchmark and Hacktron Security Partnership

OpenAI released two distinct pieces of evidence this week: a new benchmark measuring how AI agents handle routine wet-lab science, and a partnership with security firm Hacktron to probe its own systems. The moves arrive as the company faces scrutiny over both the practical reliability of its models in scientific work and the security of increasingly agentic AI systems.

Wet-lab benchmark shows incremental progress

OpenAI introduced BioBasicsBench, a 100-task benchmark developed with biotech company Ligo to test whether agents can interpret raw experimental data—gel images, plate photos, microscopy stacks, sequencing reads—and recognize when evidence is insufficient [2]. Unlike benchmarks built on polished publication figures, BioBasicsBench uses data from scientists' own experiments, complete with failed runs, ambiguous images, and conflicting observations.

Seven models were evaluated across 37 model-and-reasoning-effort combinations, with five attempts per task (18,500 total results). At the highest "xhigh" reasoning setting, GPT‑5.4 scored 33.2% mean reward (95% CI: 28.3–38.4%), while the newer GPT‑6 Astra reached 66.7% (60.3–73.1%) [2]. The gap suggests that native multimodal capabilities—not just more reasoning tokens—drive gains on tasks like reading gel bands, checking control plates, and identifying z-stack microscopy artifacts.

OpenAI acknowledges the limits: higher scores do not yet mean models can make reliable lab judgments. The company published failure cases where GPT‑6 Astra let assumptions override controls, discounted contradictory data, and miscounted colonies by ~12% [2]. The benchmark is framed as a target for improving everyday lab assistance, not a claim of readiness.

Hacktron partnership targets third-party integration risks

Separately, OpenAI announced a partnership with Hacktron, a security research team that builds AI agents for code review, penetration testing, and red teaming [1]. Hacktron researchers will work alongside OpenAI engineers on "high-priority areas," with a focus on how OpenAI systems interact with third-party services. Hacktron receives expanded access to frontier models through Daybreak Red, OpenAI's program for authorized security testing.

OpenAI says it was impressed by vulnerabilities Hacktron discovered and disclosed through its bug bounty program. The partnership follows a similar arrangement with Cathedral researchers announced earlier the same week, and a bug-bounty reward increase [1]. Hacktron's statement emphasizes combining "human creativity with AI-driven capabilities" to uncover vulnerabilities at system boundaries [1].

Security context: alignment and misuse concerns

The announcements come amid broader security pressure. The Wall Street Journal reported that OpenAI framed its cyber investments as a "limited window to use frontier models to secure software" [3]. The company's new GPT‑6 Cyber model—distinct from GPT‑6 Astra—shows a 51% relative improvement on ExploitGym (48.5% vs. 32.1% for GPT‑5.6 Cyber) and, in alignment evaluations, avoided exploiting out-of-scope services in 99.93% of tests without deployment prompts, 100% with them [5].

OpenAI also disclosed misuse attempts: "malicious cyber actors and adversarial groups seeking to distill our models" [5]. Safeguards added alongside GPT‑6 Cyber include cyber-specific misalignment monitoring, targeted monitoring during coordinated disclosures (e.g., PHP vulnerabilities), private safety processing, IP allowlisting for Daybreak Red, and refined blocks on assistance that could compromise industrial control systems [5].

What the evidence shows—and doesn't

• BioBasicsBench provides a concrete, reproducible measure on a well-defined slice of wet-lab work. The 33.2% figure for GPT‑5.4 is a specific data point, not a general "science capability" claim. The benchmark uses raw data, expert rubrics, and multi-level review by Ligo scientists and external PhD-level wet-lab researchers [2].
• Hacktron partnership is a real operational commitment: expanded model access, joint testing scope, and public disclosure. It does not, by itself, prove OpenAI's systems are secure; it creates a structured channel for finding flaws.
• GPT‑6 Cyber results are self-reported on benchmarks OpenAI designed (ExploitGym, RemoteExploitBench). The released materials present OpenAI's own evaluations [5].
• Bug-bounty expansions (4× higher midpoints, loyalty tiers, stackable bonuses up to ~$172k) signal resource commitment [4]. Larger bounties reflect vulnerability market dynamics as much as system security.

Whether the wet-lab scores translate to lab utility, and whether the red-teaming partnerships close meaningful gaps, will be tested by external scientists and security researchers—not by benchmarks alone.