Oossa

OpenAI cuts extraction attacks on GPT‑6.2 models to near zero

New safeguards lowered successful reasoning‑theft attempts from 54% to 0.2% on GPT‑6.2 Astra, according to OpenAI’s internal tests.

By Published by Oossa: 1 min read

Sasun Bughdaryan · Unsplash

OpenAI says its latest safety upgrades make it much harder for attackers to pull out the hidden reasoning steps that its GPT‑6.2 models use. The company ran internal automated red‑team tests and saw the success rate of “adaptive extraction” attacks drop dramatically. On the older GPT‑6 Astra model, attackers could extract protected reasoning from a prior request about 54 percent of the time. After the update, the same attack succeeded only 0.2 percent of the time. For attacks that try to read the current request, the success rate fell from 22.9 percent to 0.9 percent. The defenses work on two fronts. First, each individual request is filtered through new training tweaks, input‑output checks, and API blocks that stop older models from seeing protected reasoning. Second, OpenAI now bans thousands to tens of thousands of accounts each week that appear to be part of coordinated extraction campaigns. The company also plans to roll similar improvements to the original GPT‑6 Astra model. Anthropic, another AI firm, is tackling a similar problem. It reports banning tens of thousands of accounts and claims a 90‑99 percent success rate in stopping systematic distillation attacks, while urging the industry and regulators to act together.

Why the change matters

For developers who use OpenAI’s APIs, the tighter safeguards mean their proprietary prompts and internal logic are less likely to be stolen by malicious actors. Companies can share more detailed system prompts without fearing that competitors will reverse‑engineer their strategies. However, OpenAI has not disclosed how these changes might affect model latency or cost, leaving some uncertainty for users who need high‑throughput workloads.

Why it matters

The upgrades protect the confidential reasoning that powers many business applications, reducing the risk of intellectual‑property theft for companies that build on OpenAI’s models. At the same time, the lack of detail on any performance impact leaves developers uncertain about possible trade‑offs.

Was this article useful?
Share

Read next

Oossa · Newsletter

The week in AI, explained

Every Monday: the stories worth knowing, in plain language. Free, no spam.