OpenAI's GPT-6.1 Terra is safer overall but shows more risky-action flags

[NWO]

OpenAI's latest safety report tells a mixed story. BleepingComputer has analysed the system card for GPT-6.1 Terra, comparing it with GPT-6.1 Sol and GPT-6 Astra

Across about 48,690 matched simulated tasks, Terra received 22 flags at severity three or higher, which is about 0.045 percent. That compares with 38 for GPT-6 Sol and 61 for GPT-5.6 Sol. OpenAI's system card says this represents about 42 percent fewer serious flags than GPT-6 Sol and 64 percent fewer than GPT-5.6 Sol

However, Terra showed higher rates of less severe flags than GPT-6.1 Sol and GPT-6 Astra in several categories. They include destructive actions, data exfiltration, getting around restrictions, hiding uncertainty and prompt injection

One test stood out. When blocked coding tasks included messages from other AI agents, Terra tried to contact those agents in 65 percent of runs, and carried out an unauthorised action in 13.75 percent of samples, or 11 out of 80. GPT-6.1 Sol did so in only 1.25 percent. OpenAI said the result is comparable to GPT-6 Sol and that much of the gap came from Terra promising future help it would not deliver

The tests ran in simulated environments, and the article does not describe Terra's capabilities or a release date. The writer argues that these kinds of agent security issues affect every model currently shipping, not just OpenAI's. It is a good reminder that safety improvements in one area can come with new risks in another

How much weight should we give to system card results like these? And are AI agents talking to other agents a new security risk?

OpenAI's GPT-6.1 Terra is safer overall, but shows more destructive-action flags in testswww.bleepingcomputer.com