Become a fan of Slashdot on Facebook

 


Forgot your password?
Close
typodupeerror

Submission + - Elon Muskâ(TM)s xAI admits Grok AI was manipulated after South Africa genoc (betanews.com)

Submitted by BrianFagioli
BrianFagioli writes: xAI is under fire after it revealed an employee tampered with Grok, its controversial AI bot on X, forcing it to make a statement about the South African genocide situation. According to xAI, the incident happened on May 14 at around 3:15 AM PST. An employee made an unauthorized change to Grokâ(TM)s prompt, instructing it to deliver a predetermined answer on a sensitive political issue. This goes against the companyâ(TM)s stated values and policies.

The company admits the prompt change somehow skipped its normal code review process. xAI now says itâ(TM)s tightening up its procedures. From now on, every Grok system prompt will be published publicly on GitHub. You can access them here. Anyone can look at them and even give feedback. xAI wants this transparency to help win back some trust in Grok as a so-called truth-seeking AI.

To prevent something like this from happening again, xAI says it will also add extra steps to its internal code review, making it harder for staff to sneak in changes without approval. On top of that, the company is putting together a 24/7 monitoring team. This group will be ready to react if Grok gives another questionable answer that the automatic systems miss.

Comment Re:Now... (Score 1) 85

by hawk (#65379507) Attached to: Judge Slams Lawyers For 'Bogus AI-Generated Research'

the lawyers on the other side should be looking up these surprising "cases" they didn't find.

At that point, they file a motion to set aside.

And there is *no* time limit und Rule 60(d)(3) for "fraud upon the court."

Once this is filed, and the judge scraped from the ceiling, there will be a hearing so fast as to make heads spin!

hawk, esq.

Comment Huge difference (Score 2, Informative) 131

and comparing resistance to early reactions against spell-check

With spell check, you knew you were receiving the correct spelling. With ChatGPT, or other LLMs, you can, and will, be told things which aren't remotely true. For example, Google's AI said to use glue to hold cheese on your pizza.

Recently, that thing called Grok has been deliberately programmed to spew nonsensical bullshit about "genocide" in South African farmers REGARDLESS of the topic being discussed.

Until things get worked out, comparing ChatGPT, or any LLM, to spell check is a nonsensical comparison.

Submission + - Palantir targeted WikiLeaks

Submitted by Anonymous Coward
An anonymous reader writes: The WikiLeaks Threat

“An Overview by Palantir Technologies, HBGary Federal, and Berico Technologies”
--

Speed is crucial!

– There is no time to develop an infrastructure to support this investigaton

– The threat demands a comprehensive analysis capability now

Combatng this threat requires advanced subject matter expertise in cybersecurity, insider threats, counter cyber-fraud, targeting analysis, social media exploitaton

Palantir Technologies, HBGary Federal, and Berico Technologies represent deep domain knowledge in each of these areas

– They can be deployed tomorrow against this threat as a unified and cohesive investigatve analysis cell

Feed the fuel between the feuding groups. Disinformation. Create messages around actions to sabotage or discredit the opposing organization. Submit fake documents and then call out the error.

Create concern over the security of the infrastructure. Create exposure stories. If the process is believed to not be secure they are done.

Cyber attacks against the infrastructure to get data on document submitters. This would kill the project. Since the servers are now in Sweden and France putting a team together to get access is more straighqorward.

Media campaign to push the radical and reckless nature of wikileaks activities. Sustained pressure. Does nothing for the fanatics, but creates concern and doubt amongst moderates.

Search for leaks. Use social media to profile and identify risky behavior of employees.

Comment Re:Why not state which ones? (Score 5, Insightful) 90

I'm of two minds on this. On one hand, if you identify the company, yes, people will know what to look for and can either replace the inverter or get a whole new panel.

On the other hand, by identifying the company, you've tipped them off that you know what they did and the company will try to find a different way to do the same thing.

On the other other hand, by not identifying the company this keeps China in the dark about which ones were found and allows time for companies to look for the same thing in other brands. If it turns out it's only these two brands then you can ban them from use. But if you find this commonality among a range of Chinese brands, you can use that as direct evidence of government involvement.

Submission + - Rogue Communication Devices Found In Chinese Produced Solar Panel Inverters (reuters.com) 1

Submitted by Gilmoure
Gilmoure writes: Looks like one more vector for Chinese influence on western utilities.

Reuters: "Using the rogue communication devices to skirt firewalls and switch off inverters remotely, or change their settings, could destabilise power grids, damage energy infrastructure, and trigger widespread blackouts, experts said."

Slashdot Top Deals

What's the difference between a computer salesman and a used car salesman? A used car salesman knows when he's lying.

Close