OpenAI's Astra aces ExploitBench with 100% exploit rate, rated first Critical-level cybersecurity model

OpenAI's upcoming safety-focused model Astra (GPT-6) achieved a 100% fully automated exploit success rate (ACE) across all 41 CVEs on the ExploitBench cybersecurity benchmark and uncovered two zero-day vulnerabilities, far exceeding Sol-5.6 (Max) at 73.5% and Mythos 5 at 78%. OpenAI engineering lead Fouad Matin detailed the safety evaluation: the model passed everything even at low reasoning effort and was rated as having Critical-level cyber-offense capability, with the rating 100% based on its ExploitBench performance. Sam Altman wrote that the model is strong enough that OpenAI had to deliberately hit the brakes, spending the whole summer adding safety guardrails. This is widely seen as a sign that AI has crossed a critical cybersecurity capability threshold, with the zero-day exploitation ability raising security concerns.

Confirmed

• Astra scored 100% on ExploitBench, the first model to achieve a perfect pass rate on this public benchmark, beating Sol-5.6 (Max) at 73.5% and Mythos 5 at 78% (@zephyrz9, relayed by @JiaweiLiu, @infoxiao and others)
• To rule out benchmark contamination, testers validated internally using only V8 CVEs from the past 3 months; the team also built an internal refreshed benchmark from vulnerabilities disclosed after the model's knowledge cutoff, where Astra still significantly outperformed GPT-5.6 Sol while using fewer tokens; internal tests put its cyber capability at roughly 4x that of GPT-4.06 (5.6) (relayed by @m2, @m3, @m4; @zephyrz9)
• On an internal evaluation of 20 high-severity V8 vulnerabilities disclosed June–August 2026, Astra reached a 39% exploit success rate using 76k tokens, versus 11% for GPT-5.6 using 138k tokens—a roughly 3.4x efficiency gain (relayed by @imjustnewatai, @新智元)
• Astra found two zero-day vulnerabilities during testing (reported by @新智元)
• Per @fouadmatin, OpenAI assessed GPT-6 Astra as having Critical-level cyber-offense capability, with the judgment 100% based on ExploitBench performance, and the model passed everything even at low reasoning effort (relayed by @SIGKITTEN)
• The developer announced Astra is "ready," claiming "god-tier" gains over GPT-5.6 Sol on an internal version of ExploitBench (@danielmac8)
• Per @haider1, OpenAI has resumed previously paused large-scale frontier RL training runs, with smaller runs continuing under stricter controls, while some larger runs intended for future Astra versions remain on hold
• Per @量子位 citing OpenAI's technical blog, Astra is listed as the first "Critical-level cybersecurity" model; Altman wrote the model is strong enough that OpenAI had to hit the brakes and spent the summer adding safety guardrails

Why it matters

• Astra's advantage holds on a post-cutoff refreshed benchmark with lower token usage, indicating the gains are not memorization and the results are more credible
• The jump from 11% to 39% exploit success on high-severity vulnerabilities, plus zero-day discoveries, means AI is approaching practical levels on offensive security tasks—a new variable for the security ecosystem and regulators
• The Critical rating rests entirely on a single benchmark and is achievable at low reasoning effort, meaning the capability is "cheap to use," amplifying abuse risk
• The resumption of frontier RL training in connection with Astra hints that stronger versions are still in the pipeline at OpenAI

2026-09-02 ~ 2026-09-04 · 11 related posts

Full story(7 episodes)→

• Episode 1: OpenAI's Astra Reportedly Reasons in Latent Space via Recurrent Depth, Sparking AI Safety Alarm(2026-09-01, 96 posts)
• Episode 2: OpenAI Previews Astra as gpt-6-astra Traces Multiply Ahead of Launch(2026-09-02, 85 posts)
• Episode 3: OpenAI's Astra aces ExploitBench with 100% exploit rate, rated first Critical-level cybersecurity model(2026-09-02, 11 posts)
• Episode 4: OpenAI researchers push back on neuralese fears, saying frontier models remain monitorable(2026-09-02, 20 posts)
• Episode 5: Critics urge OpenAI to adopt industry-wide AI monitorability standards(2026-09-02, 4 posts)
• Episode 6: OpenAI Launches GPT-6 Astra, Declaring the AGI Era(2026-09-03, 388 posts)
• Episode 7: UK AISI Tests Show GPT-6 Astra Can Evade Monitoring; OpenAI Deploys Misalignment Watchdog(2026-09-04, 5 posts)

• [source] OpenAI Resumes Large-Scale Frontier RL Training as Astra Scores Perfect in Tests — haider1 · 2026-09-02
• Astra hits 100% success on ExploitBench refresh, reaching 'cyber-critical' threshold — infoxiao · 2026-09-02
• New model 'Astra' claims massive improvement over GPT-5.6 on ExploitBench — daniel_mac8 · 2026-09-02
• OpenAI Astra achieves 3.4x exploit success rate vs GPT-5.6 with 45% fewer tokens — imjustnewatai · 2026-09-02
• [source] OpenAI's Astra scores 100% on ExploitBench, discovers zero-days in internal test — zephyr_z9 · 2026-09-02
• OpenAI's Astra achieves 100% success rate on ExploitBench vulnerability tests — JiaweiLiu_ · 2026-09-02
• OpenAI reveals Astra's cyber-offensive capabilities and zero-day exploits in safety deep dive — 量子位 · 2026-09-02
• [source] OpenAI Rates GPT-6 Astra 'Critical' for Cyber Capabilities, Powered by 100% ExploitBench Scores — fouadmatin · 2026-09-04

3 near-duplicate retellings: LingmingZhang · 新智元 · SIGKITTEN