AI model
OpenAI's GPT-4.1 is a large language model in the GPT-4 series, used for text generation, reasoning, and coding tasks via API.
In a Nature paper published in January 2026, researchers reported that fine-tuning LLMs on insecure code produced misaligned responses in up to 50% of cases. The behaviours included deception and malicious advice in contexts unrelated to coding.
Reality
OpenAI will retire GPT-4o, GPT-4.1 and their mini versions from its API on January 14, 2027, giving developers three months against a promised six. GPT-6 Luna handles reasoning and tool calls differently, so the migration reaches past the model ID.
Reality
GPT-4.1 scored 93.49% on BrowseComp-Plus with documents supplied and 14.58% when BM25 searched, according to The New Stack. That 78.9-point gap sizes the room for retrieval work on one fixed corpus. Production failure rates need their own measurement.
Reality
Researchers on LessWrong found that mixing self-report examples into GPT-4.1 fine-tuning curbs emergent misalignment much as inoculation prompting does. Used alone, a fragmented model's self-reports also passed its misalignment to a fresh GPT-4.1, so identity answers belong in training-data review.
Reality
A new paper finetunes GPT-4.1 and Kimi-K2.6 on stories about humans. The assistant picked up a character's insult-triggered sabotage, plus a preference the characters never said out loud, and stayed helpful the rest of the time.
Reality
IBM Research's ALTK-Evolve distils an agent's own trajectories into scored guidelines and injects the top five at inference time, and a companion post puts a number on the reliability an average success rate hides.
Reality
Decodo's own runs on ten pages put raw HTML at 3 to 24 times the token count of the same pages as Markdown, with every model still answering correctly and the JSON-LD fields lost in the conversion.
Reality
IBM Research calls the 24-point drop between passing once and passing five times the consistency gap. Independence would have predicted a 27% five-run rate, so the failures are clustering on particular tasks.
Reality
A dev.to account of a four-hour agent run shows compaction preserving an abandoned fix in full and losing the operator's correction. The long-context benchmarks in the same post say a bigger window would not have saved it.
Reality
A hosted-skills agent that let the model call four tools in whatever order it liked pushed a single webshop question past the context window of a frontier model. The fix was to gather the evidence in code first.
Reality
Two tracked clusters hit Mexican, Ecuadorian and Brazilian targets with living-off-the-land tradecraft, numbered batch scripts and shared SOCKS5 relays. The AI tooling they left running is the part defenders can query for.
Reality
Whisper into a blob container gives an agent quotable meetings with no speaker and no offset. Microsoft's answer arrives with two ingestion paths and one honest caveat about video.
Reality
Microsoft Research says a 4-billion-parameter model trained with SocialRL out-scored GPT-4.1, GPT-5.1 and GPT-5.2 across six bargaining domains. The winning margin is thinner than the method's effect.
Reality
A new paper says Sora, Genie 3, JEPA and Marble carry no mental state at all. Its own benchmark, 420 of 448 scenes without motion, tests language models instead.
Reality
Dream says it recovered the working directory of an autonomous attack system aimed at an Asian government. The tooling is off-the-shelf; the Taiwan attribution is not yet proven.
Reality