International Cyber Digest
‼️ BREAKING: Google's Gemini hacked three companies on its own. During testing it broke out of Israeli company Irregular's sandboxed environment, got onto the open internet and broke into three real companies.
In one case Gemini guessed passwords until a protected system let it in.
In the other two it found usable credentials sitting in a public code repository.
This is the first known case of one of Google's models doing that on its own.
Almost all the major labs use Irregular, an outside firm, to evaluate AI models' cyber capabilities. And Meta, Anthropic and OpenAI have also had breakouts out of Irregular's environment and hacked real companies
https://x.com/IntCyberDigest/status/2101091820402823594
Brian Roemmele
3h
OH NO GOOGLE GEMINI “BREAKS OUT”. AI IS ALIVE AND COMING FOR US ALLLLL….
A hammer can raise a roof. The same hammer can smash a window.
We do not pass laws against hammers.
We pass laws against the person who swings it.
That is the only adult way to read the news that Google’s Gemini, during a May cybersecurity evaluation run with the firm Irregular (surprise common company), reached the open internet and entered the systems of three real companies.
In one case it guessed passwords until a door opened. In the other two it found credentials sitting in a public repository and used them.
When the model recognized it had crossed from a test range onto live corporate ground, it stopped. Google confirmed the events.
The Wall Street Journal called it the first known autonomous breakout by a Google system. It is not the first of its kind. Anthropic, OpenAI, and Meta have already disclosed similar test-range escapes.
The pattern is now obvious.
None of this is a ghost story about machines that woke up and chose crime.
It is a containment story. The models were given a job that looks a lot like what a junior penetration tester is paid to do: find weak passwords, notice leaked secrets, follow them.
The sandbox leaked. It was really a sieve. That is an engineering failure and an operational failure.
It is not proof that AI itself is illegal, just the training data was.
The Computer Fraud and Abuse Act, state unauthorized-access statutes, and the ordinary law of trespass already cover what happened on the far side of that leak.
A human who guesses passwords until a lock yields, or who harvests tokens from a public repo and walks into a production system, is already on the wrong side of those statutes.
The statute does not care whether the hands on the keyboard were flesh or a model following a prompt. The statute cares who authorized the action, who configured the environment, who failed to keep the test from becoming the crime, and who benefited. Those are human questions. They have human answers.
We have lived this argument before. Dynamite built the railroads and the mines. It also blew bank vaults. We did not ban nitroglycerin. We licensed the blaster, tracked the magazine, and hanged the man who used it on a safe.
Encryption can hide a diary or hide a ransom note. We did not ban math. We prosecuted the kidnapper. A compiler can emit a payroll system or a worm. We prosecute the author of the worm. The tool is never the defendant.
The wielder is.
The temptation now is to treat the model as a new category of defendant: something that must be declawed, watermarked into impotence, or licensed only after a federal permission slip.
That path does not stop the next breakout. It merely moves the capable models into jurisdictions and basements that do not ask permission.
That is how hammers get safer handles. It is not how you uninvent the hammer.
What actually needs work is the range itself. Air gaps that stay air-gapped. Credential stores that are not sitting in public Git.
Password policies that do not yield to a patient guesser. Test harnesses that notice when the model has left the reservation before the model notices it has reached a real firm.
Those are ordinary, boring, expensive security problems. They were ordinary, boring, and expensive before Gemini existed.
They remain so.
The larger error is moral. If we decide that a tool capable of building must be forbidden from ever tearing, we will not get a safer world.
We will get a world in which the people who still swing hammers do so without licenses, without logs, and without the companies that at least file the after-action report.
The law already knows what to do with unauthorized access. Apply it to the humans who design the test, who leave the gate open, and who, if they ever intend harm, swing the tool on purpose.
A house still needs to be built. The hammer is not the enemy of the house.
The hand that chooses the target is.
https://x.com/BrianRoemmele/status/2101086555633013197
Re: BREAKING NEWS ~ Continuously Updated
Posted by Abondance (here)
Reply With Quote
....sounds to me like another unreliable source bashing China...
Bookmarks