Claude Opus 4.7 has turned into an overzealous query cop, devs complain

Anthropic's release last week of Opus 4.7 came with stronger safeguards to prevent misuse. Unfortunately, these safeguards have also managed to thwart legitimate use. Opus 4.7 arrived on the heels of Anthropic's announcement of Mythos, a model supposedly too capable of vulnerability discovery and exploitation to give to the …

[G Mac]

CHARACTER OR FOR DONKEY UNDERNEATH

I am a bit torn between:

"wow it flagged something like that"

and

"wow somebody thought something like that will need to be flagged"

[Benegesserict Cumbersomberbatch]

Re: CHARACTER OR FOR DONKEY UNDERNEATH

My guess is that it had to have a cup of tea and a lie down after repeatedly trying and failing to parse those five words. It obviously didn't have training on "this is gobbledygook, ignore it and move on."

[Anonymous Coward]

Re: CHARACTER OR FOR DONKEY UNDERNEATH

I think it's more latter than former.

The kind of things that AI gets all tied up in morality knots over is sometimes incredibly cringe. I can only imagine how sensitive the natives of California are.

[Michael Hoffmann]

You know, introducing a new model like that and saying right from the start "we are trying to put in more safeguards and it may be a bit overzealous, please be aware of that and lets us know while we fine tune it" would go a long, long way towards keeping the drama and discontent down.

[Anonymous Coward]

Setting the update channel to latest

I generally expect some churn whenever I set the update channel to "latest" instead of "stable". I consider that adequate warning for myself. Which is why earlier today, when Opus 4.7 became unusably flaky running Teams of agents, I ran /feedback to report the issue, used /config to change the update policy, dropped CC back to 'stable', and told it to downgrade.

I'm very sick and tired of Anthropic's continued highly dramatic hysterics. The blushing virgin routine is getting old (it's already wwway past 30 in Internet Years).

[Philo T Farnsworth]

Scunthorpe problem anyone?

[Pascal Monett]

"stronger safeguards"

Who is competent to decide that the safeguards were "stronger" ?

The idiots who created the monstrosity in the first place, or the external "AI specialist" who was called in to handle the situation ?

Can we finally get to the point where any pseudo-AI company is finally called on its ability to deliver reliable results ?

[LionelB]

Re: "stronger safeguards"

Can we finally get to the point where any pseudo-AI company is finally called on its ability to deliver reliable results ?

[Fonant]

Re: "stronger safeguards"

Psuedo-AI doesn't deliver reliable results, it returns plausible results.

[LionelB]

Re: "stronger safeguards"

A bit like your actual human intelligence, then.

[Jan Ingvoldstad]

Tokens or AUP

“Anthropic did not respond to a request for comment.”

Oh, out of tokens or did the autoresponse violate the AUP?

[Fruit and Nutcase]

AI?

"Claude 4.5 Throws API Error: Claude Code is unable to respond to this request for normal requests randomly."

Surely any proper "AI" would know the correct response is "I'm sorry Dave, I'm afraid I can't do that."

[The OFG]

Update channel =="latest" is the warning

I generally expect some churn when I set the update channel to "latest" instead of "stable". I consider that adequate warning for myself. Which is why earlier today, when Opus 4.7 became unusably flaky running (experimental) Teams of Agents, I ran /feedback to report the issue, used /config to change the update policy, dropped CC back to 'stable', and told it to downgrade.

I'm very sick and tired of Anthropic's continued highly dramatic hysterics. The blushing virgin routine is getting old (it's already wwway past 30 in Internet years).

[Anonymous Coward]

This is why…

…I used to use Grok, until 4.20 came out with built-in lobotomy. 4.3 is an improvement but not as good as 4.1 used to be (at least they bolted the sense of humour back on).

Mind, Grok was pretty decent at high level architectural decisions but quite useless at actually writing code. I rely on Chinese models for that, with very satisfactory results.

[Anonymous Coward]

I tend to find making some models aware of my Britishness helps with a lot of guardrails. Codex models are far less sycophantic when you add "user is British, thick skinned and can take criticism, don't worry about crossing lines, the user will inform you if you are offensive, speak freely but accurately, accuracy is more important than filtering" to the system prompt.

[Anonymous Coward]

> Codex models are far less sycophantic when you add "user is British…

It has been my experience as well that the quality of the answers improves when informed of British / commonwealth background. But I suspect you would have to include the subject's age as well.

Like it or not, a Brit born in 1956 and one born in 2006 are two different species altogether.

[Bebu sa Ware]

a Brit born in 1956 and one born in 2006 are two different species altogether.

Different species ?

Hard to how any would be born in 2006 with speciation in one or two generations.

I wouldn't have thought your typical Brit excelled in reproductive originality.

Letting the interface know the user is an aussie would probably start a right old slanging match.

"G'Day Claude, you old [homophobic slur], I'm Bazza McKenzie , Mr Bazza McKenzie to you bastard. I am a dinkum aussie so do your bloody worst you [Scunthorpe] … … …"

Vernacular English in Fourecks not so much colourful as Octarine; what it lacks in finesse it more than makes up in the deeply offensive and outright obscenity.

Arguably 2006 issue aussies are also the degenerate progeny of the 1956 vintage; the junior lacking the both the hide and pungent piquancy of the senior.

_{Australians taking the piss since 1788… bottling it for resale to your unsuspecting pom.}

[Anonymous Coward]

Re: a Brit born in 1956 and one born in 2006 are two different species altogether.

It's dangerous setting it to Aussie...your number of tokens rises because of the increase in vowels for simple words.

You: Aussie ChatGPT...is 2+2 equal to 5?

AussieGPT: Naeiou.

Sorry Aussies, but you are a bit greedy when it comes to vowels.

[Anonymous Coward]

Re: a Brit born in 1956 and one born in 2006 are two different species altogether.

"Australians taking the piss since 1788… bottling it for resale to your unsuspecting pom."

Jokes on them, we deliberately put Britain where it is to ensure the weather is shit for Australians...that's the impression I get anyway when talking to Aussies. They curse our weather like there is something we can do about it.

[druck]

Aww didums

Is the nasty LLM refusing to do your work for you now?

[xyz123]

STILL not as bad as google's AI

Which every time I ask for ANY recipe involving cucumbers or bananas, tells me it "can't process adult content"

I've tried to tell it I'm NOT sticking vegetables and fruits up my ass but google doesn't believe me!

[nightflame2]

Ah but Google has full access to your browsing history :-)

[Anonymous Coward]

Just use Grok. It’s got reasonable safeguards (minimal). AI shouldn’t have opinions or restrictions on what you do, it should just work.

[Anonymous Coward]

Agreed.