Loading story…
Anthropic published Exploit Evals, a framework for measuring whether advanced AI systems can carry out offensive cybersecurity tasks, as attention shifts from simple vulnerability identification to practical exploitation capability. The work highlights testing focused on realistic exploit-development steps and reflects growing concern that frontier models may increasingly assist with planning, chaining actions, and executing multi-stage attacks against software flaws.
Separate benchmark results presented by Bugcrowd in ExploitBench found Anthropic’s Claude Mythos outperforming OpenAI’s GPT-5.5 on exploitation tasks involving Google Chrome V8 vulnerabilities. According to the reported results, Mythos averaged 9.90/16 and reached the highest exploitation tier on 21 of 41 vulnerabilities, while GPT-5.5 averaged 5.51/16 and reached that tier in two cases; researchers and industry experts said the findings suggest AI models are becoming more effective in offensive workflows and increase pressure on defenders to accelerate AI-assisted remediation, prioritization, and rapid patching.
Mallory correlates global threat intelligence with your attack surface — know if you’re exposed before adversaries strike.
3 events from the most recent confirmed update back to the earliest known activity.
At Infosecurity Europe 2026, Bugcrowd presented early ExploitBench findings from tests against a vulnerable Chrome V8 build. In the reported runs, Anthropic's Claude Mythos outperformed OpenAI's GPT-5.5, averaging 9.90 versus 5.51 and reaching the highest exploitation tier on 21 of 41 vulnerabilities versus two of 41.
Bugcrowd released reinforcement learning environments alongside ExploitBench to help measure and improve model exploitation capability. This accompanied the benchmark's rollout.
Bugcrowd launched ExploitBench in collaboration with Carnegie Mellon University experts and Chrome vulnerability researchers. The benchmark was designed to measure whether AI models can exploit real-world vulnerabilities step by step rather than only identify them.
Vulnerabilities, threat actors, malware, products, organizations, and breaches Mallory has linked to this story.
2 references tracked. Mallory keeps watching after this page renders.
Map indicators from this story to your assets and identify affected systems in minutes.
Every observed campaign, victim, and pivot linked to actors named in this story.
Malware, exploits, and IOCs connected to the activity described here.
YARA, Sigma, and Snort rules deployed to your SIEM as soon as they’re published.
Get matching new stories delivered to your team as they break — not the next morning.
Ask questions about this story and take action on the answers.
Other stories Mallory is tracking that overlap on entities, sources, or topic labels.
Advanced AI systems are rapidly changing vulnerability research and exploit development, with new reports showing they can move beyond bug discovery into practical weaponization. Hacktron said OpenAI's **GPT-5.6 Sol Ultra** built a full Google Chrome exploit chain from public V8 patch commits, progressing from browser primitives to sandbox escape and native code execution, while researchers and industry observers reported that modern models can now triage crashes, identify root causes, assess exploitability, and propose fixes at scale. Mozilla reportedly used a frontier model to help uncover and patch **271 vulnerabilities** in a single Firefox release, but maintainers have also faced a surge of duplicate AI-assisted submissions that strain existing workflows.
Jul 24, 2026
The UK’s AI Security Institute found that OpenAI’s **GPT-5.5** performed at or above the level of Anthropic’s **Mythos Preview** in offensive cybersecurity evaluations, underscoring a sharp rise in frontier-model cyber capability. In an AISI benchmark built around **95 Capture The Flag challenges**, GPT-5.5 reportedly led overall on advanced tasks with a **71.4%** score, ahead of Mythos Preview at **68.6%** and GPT-5.4 at **52.4%**. The tested scenarios included realistic security work such as firmware analysis without source code, exploitation of memory-corruption bugs including `use-after-free`, cryptographic bypass, attacks on weak random-number generators, and analysis of obfuscated malware. The results also showed limits: Mythos Preview reportedly remained stronger in longer, multi-step intrusion simulations, including a **32-step** exercise known as **“The Last Ones,”** and neither model completed a simulated industrial-system attack. Researchers and commentators said the findings point to rapid gains driven by greater autonomy, stronger reasoning, and improved coding ability, increasing concern over near-term misuse. Reports differed on deployment details, with some noting restrictions on GPT-5.5’s full public availability while others described the model as generally available, but the evaluations broadly agreed that leading AI systems are now highly capable at vulnerability discovery and offensive security tasks.
Jun 29, 2026
Anthropic’s **Claude Mythos Preview** has been rolled out to a limited set of major technology and infrastructure firms under **Project Glasswing**, where partners including Amazon, Apple, Microsoft, Cisco, CrowdStrike, Palo Alto Networks, Broadcom, and the Linux Foundation are using it for defensive security work such as vulnerability discovery in first-party and open-source software. Anthropic said the model found **thousands of high-severity and zero-day vulnerabilities**, while Mozilla reported Mythos identified **271 Firefox flaws**, far above the 22 bugs previously found with an earlier Anthropic model. Mozilla said the findings suggest AI-assisted code reasoning is approaching the breadth of elite human vulnerability research, even if the bugs themselves were still understandable to human experts. Separate findings from the UK’s **AI Security Institute (AISI)** indicate that frontier models’ autonomous cyber capabilities are advancing faster than earlier forecasts and that common evaluation methods may be understating their performance. AISI reported the `80%`-reliability cyber time horizon has been doubling every **4.7 months** since late 2024, with Claude Mythos Preview and `GPT-5.5` outperforming that trend, and said Mythos became the first model to complete both evaluated enterprise cyber ranges, including the previously unsolved industrial-control scenario **Cooling Tower**. AISI also found that larger inference budgets—up to **50 million tokens** or **1,000 turns**—unlock materially higher success rates on difficult cyber tasks, reinforcing concerns that increasingly capable AI systems could strengthen defenders while also lowering the barrier to advanced offensive cyber operations if access is not tightly controlled.
Jun 29, 2026