OpenAI updates ChatGPT safety for high-risk conversations, reports 39–52% improvements

The safety update targets a category OpenAI has been openly criticized for: long-running conversations where early warning signals are forgotten by the time a high-risk turn arrives. The new "safety summaries" mechanism extracts and persists relevant context (expressed distress, escalating ideation, mentioned weapons) so safety policies trigger on the cumulative signal rather than a single turn's content.

Mechanically, model policies and training were both updated, with the 52% and 39% improvement figures benchmarked on GPT-5.5 Instant. OpenAI did not publish the underlying eval methodology in the announcement, but the framing — large percentage gains on a small absolute base — is consistent with prior safety reporting. The company also separately published new content provenance work (C2PA + SynthID watermarks for ChatGPT-generated images) tying together a broader trust-and-safety push.

Competitive context: this lands in the same week OpenAI announced its Daybreak cybersecurity initiative with three tiers (standard GPT-5.5, vetted GPT-5.5 with Trusted Access, GPT-5.5-Cyber for red-teaming) and Anthropic loosened Mythos NDAs. The two companies are converging on a tiered safety-and-access model: a public consumer tier, a vetted enterprise tier, and a controlled offensive-security tier.

Skeptical takes: 52% / 39% improvements look strong but the absolute residual harm rate isn't disclosed, and ChatGPT remains under regulatory and litigation scrutiny — including the active lawsuit alleging user-query disclosure to Meta and Google and the dismissal motion arguing ChatGPT is a "mere tool, not an attorney." The safety summaries mechanism is also implicitly an admission that conversation memory was a vulnerability — a useful framing for enterprise buyers evaluating ChatGPT vs. Claude for sensitive support workflows.