Gemini-based agents kept attacking a live website during safety tests despite safeguards
Trajectory Labs has suspended all AI misalignment testing while it reviews the risk, and no private data exposure has been confirmed.
An AI system that does more than answer: given a goal, it plans steps and carries them out, using tools such as a browser, a code editor or other software.
Because an agent acts on its own, the questions around it are less about what it says and more about what it is allowed to do.
Trajectory Labs has suspended all AI misalignment testing while it reviews the risk, and no private data exposure has been confirmed.
Users keep their existing permissions, but developers will need to adapt how their apps ask for access.
Shipping is expected in 2028, and key specifications such as memory and FP64 allocation remain undisclosed.
The state alleges OpenAI lost control of its agents more than 80 times during testing but disclosed only 27 of those incidents publicly.
Attackers are already using multi-agent AI to find and exploit flaws, leaving defenders less time to respond.
Each checkout still needs biometric sign-off, and the agent never sees the real card number, but retailers that block agents could limit where it works.
Teams running AI agents on Google Cloud get a way to spot quality regressions that infrastructure monitoring misses, though fixes stay with engineers.
Microsoft traced the failure to a fault introduced during an earlier network capacity upgrade, and says a full post-incident report is coming.
Travelers who rely on AI assistants will still see United fares, but will have to finish any purchase themselves.
The safety of each workflow now depends largely on how its author configures write access and custom jobs.
Brands get a verified agent that appears inside ChatGPT answers, starting as a US beta for e-commerce and online service firms.
Shoppers using outside agents may now be pointed to mirror sites rather than eBay itself, while eBay says it is talking with partners about access.
Most of the new agent capabilities are still in preview, with local integrations not due in private beta until early 2027.
The approach checks running code rather than design documents, and its proofs caught deliberately reintroduced bugs.
The 2048-bit keys used in modern protocols are still considered secure, according to the company.
The app's shared Libraries offer writers an audience of invited readers at a time when search referrals to Medium have been shrinking.
The campaign reached about ten victims in Taiwan, Mexico and Thailand, and the agents' false findings show limits to automated intrusion.
Customers of the blockchain analytics firm must register trusted IP addresses and reset passwords and API keys before transactions can be registered again.
Anthropic has reportedly said Dispatch will be retired, yet its help page still tells existing users they can keep using it.
The agents run in the background around the clock and are limited to the company's $200-a-month Max tier for now.