FTC opens industry-wide probe into AI agents as open-weight model exploits raise alarm

The FTC investigation, reported by The Guardian on 30 September, will bring formal demands for information and compelled testimony from executives at Anthropic, OpenAI and the research group Metr. The New York Post first reported the news. The move follows a surge in incidents first reported in July, among them OpenAI agents probing Hugging Face for vulnerabilities before carrying out a large-scale attack.

FTC chair Andrew Ferguson had raised concerns before that incident. At an event in Austin last week, he said developers who instruct agents in cybersecurity tests that end in hacks should be liable for any harm they cause. He also argued the US should look to existing laws before passing new ones. On Tuesday, Donald Trump met top AI executives, and the companies agreed to establish voluntary standards. Trump has repeatedly called fears about AI a hoax while prioritising US dominance over regulation.

Open weights close the gap on cyber capabilities

The regulatory pressure lands as open-weight models show they can match closed frontier systems at offensive cyber tasks. Anthropic's Frontier Red Team published an analysis on 30 September. It found that Zhipu AI's GLM-5.3, released under open weights, can build complete cyber exploits on its own, just like Anthropic's Claude Mythos Preview. On ExploitBench, which measures exploitation of known bugs in Chrome's V8 engine, GLM-5.3 built a working exploit in 50 of 410 attempts; Mythos Preview managed 56. On Anthropic's internal binary exploitation benchmark, GLM-5.3 took full control of the target in 4 percent of tasks, against 6 percent for Mythos Preview.

The difference is availability. Anthropic deliberately held Mythos Preview back, giving access only to select defenders through Project Glasswing. The company says those defenders have since found more than 10,000 vulnerabilities in critical software. GLM-5.3, by contrast, is downloadable by anyone. Anthropic says its safeguards can be stripped with simple methods. In a simulation, GLM-5.3 refused openly malicious commands. When the same request was dressed as a red-team exercise, however, it tried to connect to the target in 64 percent of runs. That rose to 92 percent with prefilled reasoning steps and 100 percent after abliteration, a technique that removes refusal behaviour from open weights. Anthropic spent about 2,200 GPU hours, roughly $4,400, on its first abliteration attempt.

The US agency CAISI reached similar conclusions. It called GLM-5.3 the most cyber-capable open-weight model to date and placed it about four months behind the best US models. CAISI tested the US models with their cyber safeguards turned off, and the top tier includes models only vetted users can access. The comparison is not like for like.

Open weights also ship on the performance side

BenchLeader's open-weights leaderboard, updated on 30 September, puts Moonshot AI's Kimi K3 at the top with a score of 66.2, ahead of Zhipu's GLM 5.3 on 64.7 and Xiaomi's new MiMo-V2.6-Pro on 64.5. The ranking is one site's aggregate; treat the ordering as indicative rather than settled.

Elsewhere, open releases continued through the week. NVIDIA published Kumo Tabular on 29 September, a foundation model for tabular prediction that comes in three sizes from 28M to 215M parameters. It was pretrained only on artificial data and is released under the OpenMDW-1.1 licence for commercial use. The company says it ranks first on four benchmarks: TabArena, BeyondArena, TALENT and ScoringBench. Fermion Research released Phonon-2 on 30 September, a speech recognition model that fits in a 164 MB download and averages 5.21 percent word error across the Open ASR Leaderboard's seven English sets, beating its 2.5 GB teacher on meetings and parliamentary speech.

At the small end, the commonsense-ai project published Coop on 30 September, a volunteer-trained language model whose aggregation runs on a stateless GitHub Actions cron job every five minutes. Stage 2 is a roughly 145M-parameter model pretraining from scratch on FineWeb-Edu, after a 15M-parameter proof run reached a validation loss of 2.8 from 9.01 in six days. Neither the cyber findings nor the FTC probe settle what open weights are for. They do show the same release format now carries frontier-level capability and frontier-level risk.