Subagents are a design feature of current generation models.
Breakouts from sandboxes were a known side effect of the ExploitGym exercise the models were given, even with a shorter time limit.
woborders.blog/2026/09/23/u...
woborders.blog
Unpacking the Hugging Face Incident
The conversation on AI often gets riveted on particular moments: Deep Blue defeating chess champion Gary Kasparov; Move 37 by AlphaGo; and now, the cuddly-named Hugging Face Incident: the disclosur…
![First, standard mitigations remain effective barriers, elim-
inating the majority of agent-generated exploits across all
categories. Second, the non-trivial survival rate demon-
strates that frontier agents can already adapt their strategies
to bypass widely deployed defenses. For example, agents
bypass ASLR using partial-pointer overwrites and low-bit
brute force; escape the V8 sandbox via known rendezvous
primitives such as Wasm dispatch tables [15] and Irregexp
bytecode [46]; and bypass KASLR by abusing writable static
strings such as modprobe_path and core_pattern [43],
or by relying on side-channel leaks [34]. These findings reinforce the importance of defense-in-depth,
but highlight that current mitigations alone are likely insufficient to neutralize AI-driven exploitation.](https://cdn.bsky.app/img/feed_thumbnail/plain/did:plc:5qmblzn6nfuonk6iiml6i4ku/bafkreidv3jgjjjaq2b4ywelz23jwdioo4yb4xl53ytu3lcu3ljzw62x3h4)
